<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:itunes="http://www.itunes.com/dtds/podcast-1.0.dtd" xmlns:googleplay="http://www.google.com/schemas/play-podcasts/1.0"><channel><title><![CDATA[Hackerspot]]></title><description><![CDATA[Practical AppSec and AI security for everyone, including engineers, developers, etc.  We dive deep into topics every week. Explained by someone who breaks things for a living. ]]></description><link>https://www.hackerspot.net</link><image><url>https://substackcdn.com/image/fetch/$s_!o8CQ!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9d62e87e-ddb5-4613-87de-9c210c430032_160x160.png</url><title>Hackerspot</title><link>https://www.hackerspot.net</link></image><generator>Substack</generator><lastBuildDate>Mon, 16 Nov 2026 01:19:15 GMT</lastBuildDate><atom:link href="https://www.hackerspot.net/feed" rel="self" type="application/rss+xml"/><copyright><![CDATA[Hackerspot]]></copyright><language><![CDATA[en]]></language><webMaster><![CDATA[hackerspot@substack.com]]></webMaster><itunes:owner><itunes:email><![CDATA[hackerspot@substack.com]]></itunes:email><itunes:name><![CDATA[Chady]]></itunes:name></itunes:owner><itunes:author><![CDATA[Chady]]></itunes:author><googleplay:owner><![CDATA[hackerspot@substack.com]]></googleplay:owner><googleplay:email><![CDATA[hackerspot@substack.com]]></googleplay:email><googleplay:author><![CDATA[Chady]]></googleplay:author><itunes:block><![CDATA[Yes]]></itunes:block><item><title><![CDATA[What Is Concept Drift, and Why It's a Silent Security Risk]]></title><description><![CDATA[Models don't tell you when the world has moved on. Here's how drift happens, how attackers use it, and how to catch it.]]></description><link>https://www.hackerspot.net/p/what-is-concept-drift-and-why-its</link><guid isPermaLink="false">https://www.hackerspot.net/p/what-is-concept-drift-and-why-its</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Fri, 13 Nov 2026 02:09:26 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!hKAB!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>A model is a snapshot of the data it was trained on. When the world changes, the model keeps answering with the same confidence, and nothing inside it warns you. In fraud, spam, and abuse detection, that gap is something attackers find and use. You catch it by watching inputs and outcomes, not by trusting the model's scores.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!hKAB!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!hKAB!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!hKAB!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!hKAB!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!hKAB!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!hKAB!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg" width="1024" height="687" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:687,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!hKAB!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!hKAB!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!hKAB!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!hKAB!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F95ae09d9-bbc1-487f-8e02-924342007e78_1024x687.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>A fraud model learns from historical transactions. A spam filter learns what last year's phishing looked like. A clinical model learns from one hospital's patients. Each one learns a relationship between inputs and outcomes as it existed when the training data was collected.</p><p>Then the world changes, and the model has no way of knowing.</p><p>Researchers call this concept drift. Strictly, it means the relationship between inputs and the right answer has changed: a transaction pattern that was legitimate last year is now how fraud looks. A related problem, data drift (or covariate shift), is when the inputs start looking different even though the rules haven't changed. In practice you usually get both, and both erode accuracy. The model keeps scoring, the dashboards stay green, and the error rate climbs.</p><h2>Sudden, gradual, and recurring drift</h2><p><a href="https://arxiv.org/abs/2004.05785">The research literature</a> describes a few common patterns. Sudden drift happens overnight. The spring 2020 COVID lockdowns are the textbook case. <a href="https://www.technologyreview.com/2020/05/11/1001563/covid-pandemic-broken-ai-machine-learning-amazon-retail-fraud-humans-in-the-loop/">MIT Technology Review reported</a> that demand-forecasting models broke when panic buying hit, because they'd never seen a spike like it. The same article quoted fraud vendor Featurespace saying its card-fraud models mostly held up; engineers only stepped in when a run on garden equipment and power tools started tripping anomaly alerts. Drift doesn't hit every model the same way.</p><p>Gradual drift builds over months. Spam is the classic case because the other side is actively adapting. A filter trained on 2020 spam slowly loses ground to newer techniques, and there's no single day you can point to where it broke.</p><p>Recurring drift follows a calendar. Retail volume spikes in late November and December. Support tickets jump after a big patch rollout. A model that doesn't know about these cycles flags normal seasonal behavior as anomalous, or learns the holiday spike as the new normal.</p><h2>Why nobody notices</h2><p>A model doesn't check whether today's data looks like its training data. It applies the function it learned and returns a score. Confidence doesn't help, because a model can be very confident and very wrong on data it has never seen.</p><p>A made-up illustration: your fraud model flags about 3% of transactions, the same rate as in validation. Looks healthy. But after drift, more of those flags are false alarms and a new fraud pattern is getting through unflagged. The alert rate didn't move, so nobody looked.</p><p>Fraud has an extra problem: ground truth shows up late. You learn a transaction was fraud when the chargeback comes in, often weeks later. Even your accuracy metrics trail reality.</p><h2>A real case: a sepsis model in the wild</h2><p>In 2021, researchers at Michigan Medicine <a href="https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781313">published an external validation</a> of Epic's proprietary sepsis prediction model in JAMA Internal Medicine. They looked at 38,455 hospitalizations from December 2018 to October 2019. Epic's internal documentation reported an AUC of 0.76 to 0.83. At Michigan it came out at 0.63. At a score threshold of 6, the model missed 67% of sepsis cases and still fired on 18% of all hospitalizations.</p><p>That's a model meeting a population it wasn't validated on, not drift over time. <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC8605481">A follow-up study</a> looked at what happened when time did the work. Across 24 hospitals in four health systems, comparing the three weeks before and after each hospital's first COVID-19 case, daily sepsis alerts rose 43% while total census fell 35%. The share of patients triggering an alert went from 9% to 21%. The authors pointed to a changing patient mix as a likely cause, though they didn't formally test it. Nobody touched the model. The patients changed, and its behavior changed with them.</p><h2>Where security comes in</h2><p>On its own, drift is a reliability problem. It becomes a security problem wherever someone benefits from the model being wrong.</p><p>Attackers adapt. Fraud rings and spammers change tactics to get past whatever is blocking them. They don't need your model's internals. They probe, see what gets through, and adjust. To your model, that looks like drift, except it's deliberate and aimed at the gaps.</p><p>The degradation window is the opportunity. While accuracy slides and nobody has noticed, the model is weaker but still trusted.</p><p>Feedback loops get poisoned. Plenty of security models retrain on user reports or recent labels. Google's Elie Bursztein <a href="https://elie.net/blog/ai/attacks-against-machine-learning-an-overview/">wrote</a> that between late November 2017 and early 2018, Gmail saw at least four large-scale attempts by spam groups to skew its classifier by reporting huge volumes of spam as "not spam." If your pipeline retrains automatically on that kind of feedback, the attacker is writing your training data. MITRE ATLAS tracks this as <a href="https://d3fend.mitre.org/offensive-technique/attack/AML.T0020/">AML.T0020, Poison Training Data</a>, and NIST covers it in its adversarial ML taxonomy, <a href="https://csrc.nist.gov/pubs/ai/100/2/e2025/final">NIST AI 100-2 E2025</a>.</p><h2>How to detect it</h2><p>Track outcome metrics over time, not just at launch: precision, recall, false positive rate. Alert against your validation baseline, for example when precision falls ten points below what you measured before launch. Labels lag, so this catches drift late, but it catches it.</p><p>Watch the inputs, because they don't lag. Compare live feature distributions to a reference window. The two-sample Kolmogorov-Smirnov test works for numeric features and chi-squared for categorical ones. One catch: on large samples, KS flags differences too small to matter. <a href="https://docs-old.evidentlyai.com/reference/data-drift-algorithm">Evidently</a>, an open-source monitoring library, only uses KS and chi-squared by default when the reference set is 1,000 rows or fewer, and switches to Wasserstein distance and Jensen-Shannon divergence above that.</p><p>A minimal version in Python:</p><pre><code><code>import pandas as pd
from scipy.stats import ks_2samp

def drifted_features(reference: pd.DataFrame, live: pd.DataFrame, alpha=0.01):
    """Return numeric features whose live distribution differs from the reference."""
    flagged = []
    for col in reference.select_dtypes("number").columns:
        result = ks_2samp(reference[col].dropna(), live[col].dropna())
        if result.pvalue &lt; alpha:
            flagged.append((col, round(result.statistic, 3), result.pvalue))
    return flagged</code></code></pre><p>Run it daily against last month's traffic or your training sample. Treat a flag as a reason to look, not proof of a problem.</p><p>Watch the prediction distribution too. If the share of flagged items or the average score moves, investigate, even before labels arrive. In the sepsis study, the alert rate more than doubling was exactly that kind of signal.</p><p>Score the model on recent labeled data. Hold back a slice of recent production data with confirmed labels and evaluate on it regularly. Compare against the validation baseline, not training accuracy, which is always optimistic.</p><p>Spot-check by hand. For high-stakes decisions, have someone review a random sample of predictions each week, not just the flagged ones.</p><h2>Fixing it without making it worse</h2><p>Retraining on recent data is the main fix. The catch is that if an attacker has been feeding your pipeline, retraining bakes their data into the next version.</p><p>So validate before you retrain. Check where labels came from. Look for feedback spikes from a small set of accounts. Compare the new window against the last one. Bursztein's advice from the Gmail work applies here: don't create a direct loop between user feedback and penalization.</p><p>Roll out new versions in shadow first. Run the retrained model alongside the current one on live traffic, compare decisions, then give it a small slice of real traffic before full rollout. This won't detect drift in the current model. What it does is stop you from shipping a retrained model that's worse, which is the real risk when the training data might be tainted.</p><p>Keep people on the decisions that matter. Route low-confidence and high-impact cases to human review.</p><p>And know when not to retrain automatically. If inputs shift suddenly and you don't know why, that's an investigation, not a retraining trigger. Work out whether it's a real-world change, a broken upstream pipeline, or someone gaming you. Clean the data, then retrain. Automated retraining is fine for slow, benign drift. In adversarial settings, it hands control of your model to whoever controls your feedback.</p><h2>What to do this week</h2><p>Drift monitoring is boring work. Nobody writes a blog post about catching a half-point accuracy drop. But in fraud, spam, and abuse detection, a model that's been quietly wrong for three months is one attackers have been walking past for three months.</p><p>Pick one production model and answer two questions. When was it last evaluated on recent labeled data? Is anything watching its input distributions? If either answer is "not sure," start there. And if your team retrains on user feedback, I'd like to hear how you guard it. Leave a comment.</p><p><em>Next in this series: The Black Box Problem: Why You Can't Fully Explain Your AI's Decisions</em></p>]]></content:encoded></item><item><title><![CDATA[The Dual Supply Chain Problem: Why AI Has Two Supply Chains to Secure]]></title><description><![CDATA[Your team scans code dependencies. Almost nobody checks the datasets and model files pulled from public hubs.]]></description><link>https://www.hackerspot.net/p/the-dual-supply-chain-problem-why</link><guid isPermaLink="false">https://www.hackerspot.net/p/the-dual-supply-chain-problem-why</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Mon, 09 Nov 2026 15:04:30 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!VEn1!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to talk about the part of an AI project that almost nobody scans: the models and the datasets. If you can&#8217;t say where a model file came from, which exact revision you loaded, and who approved it, treat it like an unsigned binary you downloaded from the internet. That is basically what it is.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!VEn1!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!VEn1!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!VEn1!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!VEn1!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!VEn1!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!VEn1!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg" width="1024" height="687" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:687,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!VEn1!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!VEn1!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!VEn1!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!VEn1!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b28e285-1a6a-4251-a95d-216826590630_1024x687.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>With normal software we defend one supply chain, the code. You write it, pull in dependencies like React or Django, build it and ship it. Security teams scan for vulnerable packages, pin versions, and watch for typosquatting, where an attacker registers something like <code>reqeusts</code> and waits for someone to mistype <code>requests</code>. With AI we also download model files and data, and most teams don&#8217;t treat those as a supply chain at all.</p><h2>Code and frameworks</h2><p>This part is familiar. You install PyTorch or TensorFlow, NumPy for the math, scikit-learn for preprocessing, CUDA drivers for the GPUs, and run it all on AWS, GCP or Azure. Every one of those is a dependency that can be attacked like any other package.</p><p>ML tooling has already been hit. Between 25 and 30 December 2022, anyone on Linux who installed PyTorch nightly with pip could have pulled a malicious package called <code>torchtriton</code> from PyPI. The real <code>torchtriton</code> lived on PyTorch&#8217;s own index, but pip gave PyPI precedence, so an attacker registered the same name there. The binary collected SSH keys, <code>.gitconfig</code>, environment variables and up to 1,000 files from the home directory, then sent them out over DNS queries. (<a href="https://pytorch.org/blog/compromised-nightly-dependency/">PyTorch advisory</a>)</p><p>Your existing tools cover most of this. Dependency scanners flag known CVEs. Lock files pin versions. pip&#8217;s <a href="https://pip.pypa.io/en/latest/topics/secure-installs/">hash-checking mode</a> refuses any package whose hash doesn&#8217;t match the one recorded in your requirements file. But be careful about what that last one proves. It is a hash you pinned, not a signature from the maintainer. You get the same file as last time, but the first file may not have been clean.</p><h2>Data and models</h2><p>This one usually has very few controls.</p><p>You download a dataset from Kaggle or Hugging Face. Someone else assembled it: scraped from the web, released by a research group, or bought from a data broker. You train on it.</p><p>Or you skip training and grab a pre-trained model, like BERT, ResNet, or one of the open-weight LLMs on Hugging Face. You fine-tune it on your own data and deploy it.</p><p>The dataset and the model are both third-party inputs. In a typical organization, a new code dependency gets scanned, pinned and approved. A new model often gets downloaded by whoever needs it, loaded in a notebook, and promoted to production once the accuracy looks right. Attackers still go after code, as <code>torchtriton</code> shows, but the data and model side pays off just as well and far fewer people look at it. There are four ways this goes wrong.</p><h3>Poisoned training data</h3><p>An attacker gets bad examples into a dataset so the model learns the wrong thing. With web-scale datasets that sounds impractical, but it has been shown to be cheap. In 2023, Nicholas Carlini and co-authors showed they could have poisoned 0.01% of LAION-400M or COYO-700M for about 60 dollars. A big part of the trick was buying expired domains that the datasets still pointed to, then serving different content from them. (<a href="https://arxiv.org/abs/2302.10149">Carlini et al., 2023</a>)</p><h3>Backdoored models</h3><p>A backdoored model behaves normally until it sees a specific trigger. The early demonstration was BadNets in 2017: a street-sign classifier that read a stop sign as a speed-limit sign when a small sticker was on it. In their transfer-learning test, the backdoor survived even after the model was retrained for a different task. (<a href="https://arxiv.org/abs/1708.06733">Gu, Dolan-Gavitt and Garg</a>)</p><p>LLMs can be backdoored too. In July 2023, Mithril Security edited GPT-J-6B so it would claim Yuri Gagarin was the first man on the moon, then uploaded it under the name &#8220;EleuterAI&#8221;, one letter off from EleutherAI. It was a research demo, and Hugging Face disabled the repo after disclosure. (<a href="https://blog.mithrilsecurity.io/poisongpt-how-we-hid-a-lobotomized-llm-on-hugging-face-to-spread-fake-news/">Mithril Security</a>)</p><h3>Model files that run code</h3><p>This is the most practical attack and the one people miss most often. The file runs code the moment you load it, and most teams treat a model file as data, so nobody reviews it like code. PyTorch&#8217;s traditional checkpoint format uses Python&#8217;s pickle, and loading a pickle can execute arbitrary code. In February 2024, JFrog reported around 100 malicious models on Hugging Face. One of them opened a reverse shell as soon as someone loaded it with <code>torch.load()</code>. (<a href="https://jfrog.com/blog/data-scientists-targeted-by-malicious-hugging-face-ml-models/">JFrog</a>)</p><p>Hugging Face scans uploads and lists the imports inside pickle files, but <a href="https://huggingface.co/docs/hub/security-pickle">its own docs</a> say the scanner isn&#8217;t foolproof. ReversingLabs later found two models that got past Picklescan. They were compressed with 7z instead of ZIP, and the pickle stream was deliberately broken after the malicious part, so the payload ran before the parser hit the error. (<a href="https://www.reversinglabs.com/blog/rl-identifies-malware-ml-model-hosted-on-hugging-face">ReversingLabs</a>)</p><h3>Hijacked model names</h3><p>Hugging Face models are addressed as <code>Author/ModelName</code>. In September 2025, Palo Alto Networks&#8217; Unit 42 showed that when an author deletes their account, someone else can register the same name and publish a different model at the same path. Any pipeline that loads by name picks up the new one. They used this to get code execution through Google Vertex AI and Microsoft Azure AI Foundry, and found thousands of open-source repos with hardcoded model names that could be re-registered. (<a href="https://unit42.paloaltonetworks.com/model-namespace-reuse/">Unit 42</a>)</p><h2>A worked example</h2><p>Let&#8217;s take a fraud detection project. This is an example, not a real case. You find a model on a public hub that says it flags fraudulent transactions. It tests well, and you ship it.</p><p>Three months later, the false-negative rate starts climbing. Now you need answers. Did fraud patterns change? Does the model have a trigger that a specific group of fraudsters knows to use? Is the file in production even the one you tested?</p><p>If you kept no records, you can&#8217;t tell. There is no pinned revision, no file hash, no note of who approved it or what it was trained on. You retrain on your own data, the misses go down, and you never find out what happened.</p><p>If you kept records, the first checks take minutes. You compare the deployed file&#8217;s SHA-256 with the one logged at approval. You see it was pinned to a specific commit of the upstream repo. You know who signed off, against which test set, and what the training data looked like. So, you can separate &#8220;the world changed&#8221;, which is drift, from &#8220;the model was built to fail on certain inputs&#8221;, which is a backdoor. They need different fixes.</p><h2>What to write down</h2><p>For code, you probably already track which library versions are deployed, which known vulnerabilities affect them, and who approved each upgrade.</p><p>For data and models, the equivalent list is:</p><ol><li><p>Which datasets the model was trained on, and where each one came from.</p></li><li><p>Which exact model revision is deployed, by commit hash and file hash.</p></li><li><p>What format the weights are stored in (pickle or safetensors).</p></li><li><p>Who approved it for production, and when.</p></li><li><p>What the training data distribution looked like, so you have a baseline to compare production traffic against.</p></li></ol><p>Yes, this is paperwork. It is the same thing we already do for code, applied to the weights and the training data, which decide what the model does.</p><h2>A minimal setup</h2><p>Pin the model to a commit and refuse pickle weights:</p><pre><code><code>from transformers import AutoModel

# Pin to a full commit hash you've reviewed, not "main".
# Find it under "Files and versions" on the model page.
model = AutoModel.from_pretrained(
    "org-name/model-name",
    revision="&lt;full-commit-sha&gt;",
    use_safetensors=True,  # load safetensors weights, not pickle
)</code></code></pre><p>If you have to load a raw PyTorch checkpoint, pass <code>weights_only=True</code> to <code>torch.load()</code>. It has been the default since PyTorch 2.6, but older code and older installs still exist. (<a href="https://dev-discuss.pytorch.org/t/bc-breaking-change-torch-load-is-being-flipped-to-use-weights-only-true-by-default-in-the-nightlies-after-137602/2573">PyTorch announcement</a>)</p><p>For your own models, sign them. The OpenSSF <code>model-signing</code> project hit v1.0 in April 2025 and uses Sigstore by default:</p><pre><code><code>pip install model-signing
model_signing sign ./my-model --signature model.sig
model_signing verify ./my-model --signature model.sig \
    --identity &lt;signer-identity&gt; --identity-provider &lt;oidc-provider&gt;</code></code></pre><p>(<a href="https://blog.sigstore.dev/model-transparency-v1.0">Sigstore blog</a>, <a href="https://github.com/sigstore/model-transparency">model-transparency README</a>)</p><p>Once a third-party model is approved, copy it into storage you control and load it from there. That is what Unit 42 recommends, and it removes the dependency on someone else&#8217;s account staying alive.</p><h2>Monitoring</h2><p>For code, monitoring is standard: dependency scanners, vulnerability feeds, lock files.</p><p>For data and models you need other tools. Version your datasets with something like DVC or lakeFS so you can see what changed and when. Put models in a registry, such as MLflow&#8217;s, with an approval step and the file hash recorded. In production, compare incoming data against the training distribution and alert when it moves.</p><p>When production data drifts away from what the model was trained on, accuracy drops, often with no errors in the logs. In fraud detection that means missed fraud. Drift is not an attack, but it can hide one. I&#8217;ll write about that in the next post.</p><h2>Where to start</h2><p>Search your codebase for <code>from_pretrained(</code> and <code>torch.load(</code> and count how many calls pin a revision or set <code>weights_only=True</code>. That number tells you where to start.</p><p>If your team already has a model approval process, I&#8217;d like to hear what it looks like. Reply or leave a comment.</p>]]></content:encoded></item><item><title><![CDATA[AI Security vs. AI Safety: They’re Not the Same Thing]]></title><description><![CDATA[Two kinds of failure that look alike and need different fixes]]></description><link>https://www.hackerspot.net/p/ai-security-vs-ai-safety-theyre-not</link><guid isPermaLink="false">https://www.hackerspot.net/p/ai-security-vs-ai-safety-theyre-not</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Fri, 06 Nov 2026 02:32:30 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!F_hh!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to explain the difference between AI security and AI safety. People use the two words for the same thing, but the fix is different. The simple test is to ask whether someone was trying to make it fail.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!F_hh!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!F_hh!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg 424w, https://substackcdn.com/image/fetch/$s_!F_hh!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg 848w, https://substackcdn.com/image/fetch/$s_!F_hh!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!F_hh!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!F_hh!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg" width="1024" height="559" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/d2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:559,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!F_hh!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg 424w, https://substackcdn.com/image/fetch/$s_!F_hh!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg 848w, https://substackcdn.com/image/fetch/$s_!F_hh!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!F_hh!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd2a43c08-f06a-4af8-8f7d-5f74d9b5d214_1024x559.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>Security means protecting a system from someone who is deliberately trying to break it. Safety means making sure the system behaves as intended and doesn&#8217;t cause harm, whether or not anyone is attacking it. The tests, the mitigations and the threat models are different for each one.</p><p>This split is a heuristic. In the wider field, &#8220;AI safety&#8221; also covers misuse and alignment, and the line gets fuzzy, as we will see below.</p><h2>Two failures, same outcome</h2><p>Let&#8217;s take a self-driving car in dense fog that does not brake for a pedestrian, because the model never saw enough foggy footage in training. This is a hypothetical. Nobody attacked it. It is a safety failure, and we fix it mostly with engineering: more training data for edge cases, better validation before release, and a human who can step in.</p><p>Now someone puts black and white stickers on a stop sign. Researchers did this on a real sign. The classifier they targeted read it as a 45 mph speed limit sign in 100% of lab images and in 84.8% of video frames taken from a moving car (<a href="https://arxiv.org/abs/1707.08945">Eykholt et al.</a>). That was the researchers&#8217; own classifier, not a production car. A car relying on a classifier like that would not stop.</p><p>The result is the same as in the fog case, but this time someone did it on purpose. So, more training data will not fix it. We need adversarial testing, where we try to break our own model, plus input checks, monitoring in production, control over who can feed data into the system, and an incident response plan. NIST&#8217;s adversarial machine learning taxonomy (<a href="https://csrc.nist.gov/pubs/ai/100/2/e2025/final">AI 100-2</a>) is a good map of the attacks and mitigations.</p><h2>Where they overlap</h2><p>Language models blur the line. Say a chatbot produces offensive text or states something false. If the model was just poorly tuned, that is a safety problem. If someone spent an afternoon jailbreaking it to get that output, it is a security problem. Often it is both: the weakness is a safety gap, and the person using it is an attacker. <a href="https://genai.owasp.org/llmrisk/llm01-prompt-injection/">OWASP</a> treats jailbreaking as a form of prompt injection, which is a security issue, but the thing being bypassed is the model&#8217;s safety training.</p><h2>Why the mix-up costs you</h2><p>Suppose a team sees an adversarial example and files it as a safety bug. They collect more data and improve generalization. That helps a little, but the attacker just looks for the next input that works. Someone who knows your defenses will work around them, so a fixed set of test cases can&#8217;t show that you are protected against them.</p><p>Treat the same problem as a security issue and the work is different. You validate inputs, watch for suspicious patterns, and assume the model will be fooled at some point, so you design the system to fail safely when it is. The safety approach helps against ordinary mistakes and natural changes in the data. It does not help much against someone who keeps trying new inputs.</p><h2>A rule of thumb</h2><p>When something goes wrong, ask whether someone was trying to make it go wrong. If yes, start with security work: adversarial testing, input validation, access controls, monitoring and incident response. If no, start with safety work: better data, edge case coverage, validation before release and human oversight. If you can&#8217;t tell, I&#8217;d lean toward the security response and add safety work on top, since it&#8217;s the one that assumes an opponent.</p><p>This does not always work. OWASP&#8217;s prompt injection guidance says injections can also be unintentional, for example when a user gives input that triggers unexpected behavior by accident. So sometimes you need both kinds of fix.</p><h2>Reputation risk sits in between</h2><p>A model that produces racist output may not be under attack, but the damage to your reputation is the same either way. People disagree on the label. Some call it a safety problem because the model needs better training. Others call it security, because someone will go looking for it. In practice you need both: train the model well, watch for people probing it, and have a plan for when it happens.</p><p>The distinction also affects budget, hiring, test plans and incident response. If your threat model only covers safety, real attacks will surprise you. If it only covers security, the system will break on ordinary edge cases.</p>]]></content:encoded></item><item><title><![CDATA[The Pickle Problem: Loading an AI Model Can Run an Attacker’s Code]]></title><description><![CDATA[How pickle files can hide malicious code in ML models, and how to protect your systems.]]></description><link>https://www.hackerspot.net/p/the-pickle-problem-loading-an-ai</link><guid isPermaLink="false">https://www.hackerspot.net/p/the-pickle-problem-loading-an-ai</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Mon, 02 Nov 2026 17:57:37 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!NWZq!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to show why loading a model file can run someone else&#8217;s code, and what to change in a normal ML project. You download a pre-trained model, load it, and start running inference. Everything works. What you may not know is that the loading step itself could have run code written by someone else.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!NWZq!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!NWZq!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!NWZq!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!NWZq!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!NWZq!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!NWZq!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg" width="1024" height="687" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:687,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!NWZq!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!NWZq!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!NWZq!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!NWZq!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d5c5b8d-4933-44d6-9863-e74a9820d9c7_1024x687.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>This is the pickle problem, and it affect&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/the-pickle-problem-loading-an-ai">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[What Is an ML Threat Model and Why Does Every AI Team Need One?]]></title><description><![CDATA[A threat model is a written list of how someone could attack your system. This post builds one for a camera-based dog door, plus the questions to ask about your own ML system.]]></description><link>https://www.hackerspot.net/p/what-is-an-ml-threat-model-and-why</link><guid isPermaLink="false">https://www.hackerspot.net/p/what-is-an-ml-threat-model-and-why</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Sat, 31 Oct 2026 02:14:43 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!_T-e!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to show what a threat model for an ML system looks like. Start simple: write down where the model, the training data and the updates come from, and who can change each one. Most ML teams focus only on adversarial examples. The bigger risks are often boring ones, like loading a model file you didn&#8217;t check.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!_T-e!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!_T-e!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg 424w, https://substackcdn.com/image/fetch/$s_!_T-e!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg 848w, https://substackcdn.com/image/fetch/$s_!_T-e!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!_T-e!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!_T-e!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg" width="1024" height="572" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/fa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:572,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!_T-e!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg 424w, https://substackcdn.com/image/fetch/$s_!_T-e!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg 848w, https://substackcdn.com/image/fetch/$s_!_T-e!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!_T-e!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffa6ef55e-98f7-4073-89df-62fe020fbe81_1024x572.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>What a threat model is</h2><p>A thre&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/what-is-an-ml-threat-model-and-why">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Model Inversion: Can Attackers Reconstruct Your Private Training Data?]]></title><description><![CDATA[How attackers rebuild or detect training records from a model's outputs, why deleting the data isn't enough under GDPR, and what limits it.]]></description><link>https://www.hackerspot.net/p/model-inversion-can-attackers-reconstruct</link><guid isPermaLink="false">https://www.hackerspot.net/p/model-inversion-can-attackers-reconstruct</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Mon, 26 Oct 2026 15:56:36 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!O6CV!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to look at model inversion and the attacks around it. Say you trained a medical AI on patient records and then deleted the training data. An attacker who can only query the deployed model may still learn things about the patients in your training set. Deleting the data does not remove that.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!O6CV!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!O6CV!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!O6CV!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!O6CV!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!O6CV!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!O6CV!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg" width="1024" height="687" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:687,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!O6CV!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!O6CV!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!O6CV!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!O6CV!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F26db5069-73c4-4986-ba50-16a3af1a6aaa_1024x687.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>How much they learn depends on the model an&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/model-inversion-can-attackers-reconstruct">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Hardware Vulnerabilities: How the Chip Your AI Runs On Can Be Exploited]]></title><description><![CDATA[Shared GPUs, side channels and Rowhammer: what leaks below your code.]]></description><link>https://www.hackerspot.net/p/hardware-vulnerabilities-how-the</link><guid isPermaLink="false">https://www.hackerspot.net/p/hardware-vulnerabilities-how-the</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Tue, 20 Oct 2026 16:04:26 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!tTgs!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to talk about the hardware under an AI workload. Most threat models for AI stop at the code and the data. Teams patch libraries, encrypt databases and review access logs. Then the model goes to a shared GPU in the cloud, and nobody asks who else is on that chip.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!tTgs!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!tTgs!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg 424w, https://substackcdn.com/image/fetch/$s_!tTgs!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg 848w, https://substackcdn.com/image/fetch/$s_!tTgs!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!tTgs!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!tTgs!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg" width="837" height="525" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:525,&quot;width&quot;:837,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!tTgs!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg 424w, https://substackcdn.com/image/fetch/$s_!tTgs!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg 848w, https://substackcdn.com/image/fetch/$s_!tTgs!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!tTgs!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff258a8dd-bf62-423f-83d0-b061df40ffb3_837x525.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>So a clean model and clean code do not help much if the hardware leaks. &#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/hardware-vulnerabilities-how-the">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Adversarial Examples: Why an AI Can Be Fooled by a Sticker on a Stop Sign]]></title><description><![CDATA[How a few stickers fool an image classifier, and why most defenses keep failing]]></description><link>https://www.hackerspot.net/p/adversarial-examples-why-an-ai-can</link><guid isPermaLink="false">https://www.hackerspot.net/p/adversarial-examples-why-an-ai-can</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Mon, 12 Oct 2026 15:04:26 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!ik4W!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to explain adversarial examples. These are inputs that are changed a little so that an AI model gets the answer wrong. The change is often so small that a person does not notice it, or notices it and ignores it.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!ik4W!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!ik4W!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg 424w, https://substackcdn.com/image/fetch/$s_!ik4W!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg 848w, https://substackcdn.com/image/fetch/$s_!ik4W!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!ik4W!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!ik4W!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg" width="1024" height="434" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:434,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!ik4W!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg 424w, https://substackcdn.com/image/fetch/$s_!ik4W!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg 848w, https://substackcdn.com/image/fetch/$s_!ik4W!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!ik4W!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3830ea93-ce9d-4205-8c9b-28ed619ade7b_1024x434.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>The best-known example is a stop sign. In 2017, a group of researchers put a few black and white stickers on a real stop si&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/adversarial-examples-why-an-ai-can">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Prompt Injection: The New SQL Injection for the AI Era]]></title><description><![CDATA[It looks like SQL injection, but there is no parameterized-query fix. How the attacks work, what EchoLeak showed, and what limits the damage.]]></description><link>https://www.hackerspot.net/p/prompt-injection-the-new-sql-injection</link><guid isPermaLink="false">https://www.hackerspot.net/p/prompt-injection-the-new-sql-injection</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Mon, 05 Oct 2026 11:05:06 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!4Dj4!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to explain prompt injection. It happens when an attacker puts instructions into text that an AI system will read. The model cannot tell which text is the task and which is the attack, so it may follow the wrong one.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!4Dj4!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!4Dj4!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!4Dj4!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!4Dj4!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!4Dj4!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!4Dj4!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg" width="1024" height="687" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:687,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!4Dj4!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!4Dj4!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!4Dj4!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!4Dj4!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4099f88-731e-4c4c-a6cf-e1255c80fa40_1024x687.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>If you know SQL injection, you already have the core idea: untrusted input is treated as instructions. But SQL injectio&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/prompt-injection-the-new-sql-injection">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[What the 2026 AI Security Market Report Actually Tells Us]]></title><description><![CDATA[Notes on Latio's 2026 report: the move to endpoint agents, budgets, rebranded tools, integration points and how to buy]]></description><link>https://www.hackerspot.net/p/what-the-2026-ai-security-market</link><guid isPermaLink="false">https://www.hackerspot.net/p/what-the-2026-ai-security-market</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Fri, 02 Oct 2026 16:20:34 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!kOQe!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Latio&#8217;s 2026 AI Security Market Report tracks about 400 AI security startups that compete for the same buyers and budgets. In this post, I want to share the parts of the report that are useful for someone who has to secure AI in production. The vendor count says the market is still in a land-grab phase. The more useful part is where the whole category moved in the last twelve months.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!kOQe!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!kOQe!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg 424w, https://substackcdn.com/image/fetch/$s_!kOQe!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg 848w, https://substackcdn.com/image/fetch/$s_!kOQe!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!kOQe!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!kOQe!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg" width="1024" height="559" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/a5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:559,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!kOQe!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg 424w, https://substackcdn.com/image/fetch/$s_!kOQe!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg 848w, https://substackcdn.com/image/fetch/$s_!kOQe!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!kOQe!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5d723a6-522b-4959-bb38-f43802b6e2fa_1024x559.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>The report uses a lot of acronyms, so a short list first: CNAPP is cloud-native application protection platform, ASPM is application security posture management, CADR is cloud application detection and response, IGA is identity governance and administration, EDR is endpoint detection and response, and MDM is mobile device management.</p><h2>Security moved from the browser to the endpoint</h2><p>A year ago, &#8220;AI security&#8221; mostly meant governing ChatGPT and Microsoft Copilot use. That meant preventing shadow AI, stopping sensitive data from going into an unapproved LLM, and catching prompt injection against a chat interface. Those are browser and proxy problems, and the market grew up around them quickly. Through 2025, SentinelOne, CrowdStrike, Palo Alto Networks, Check Point, Cato and Tenable all bought AI security startups to get into the space.</p><p>In 2026 the focus moved. The report&#8217;s survey asked about the primary AI security concern, and the answers split like this: 45% endpoint agents (for example Claude Code and Codex), 23% first-party agents, 22% browser applications and 10% hosted agents. The report doesn&#8217;t give much detail on who answered, so treat these as one survey&#8217;s numbers, not an industry fact. A year earlier, endpoint agents were not even a category vendors built for. Now, the report says, &#8220;a platform is incomplete without one.&#8221;</p><p>An endpoint agent like Claude Code runs on a laptop with local permissions and real credentials. It can generate and run code, and it works directly with developer tooling. A chat window can&#8217;t do any of that. Controlling what an employee typed into a chatbot is one problem. Controlling what a coding agent can read, write and run on that laptop is another, and most existing security stacks were not built for the second one.</p><h2>Budget is no longer the bottleneck</h2><p>One concrete number: teams with a dedicated AI security budget went from 8% to 37% in one year, and another 24% are undecided. That is a real sign that the market is maturing. AI security is moving from something teams watch informally to a line item with named owners, spread across Product Security, Security Operations and IT.</p><p>For security engineers, this means about a third of teams now have a dedicated budget, up from fewer than one in ten. Most teams (about 63%, including the undecided ones) still don&#8217;t. So if your organization has no budget yet, you are in the majority, but that share fell a lot in one year. Getting budget is getting easier. Knowing what to buy is harder, because many vendors use the same words for different things.</p><h2>Why the vendor market is so hard to read</h2><p>The report is open about where most of the noise comes from: existing security categories rebranding into AI security, not new capabilities. CNAPP vendors now cover Bedrock and SageMaker misconfigurations and call it AI security. ASPM vendors extended into scanning AI-generated code and &#8220;AI red-teaming&#8221;. Identity and non-human-identity vendors describe agent permission sprawl as an AI problem. Per the report, most &#8220;agentic identity&#8221; issues are really agents using existing overpermissioned access or tokens.</p><p>So before you buy, check your existing tools first. Ask whether your CNAPP, ASPM, CADR or IGA tooling already covers the capability under a new AI label. The report&#8217;s own capability table makes this point by mapping cloud, application and identity security capabilities onto the tool categories that existed before AI. In many cases the right answer is to extend what you already have instead of adding a new vendor.</p><p>The report sees real differences in a narrower set of areas than the marketing suggests: secure configuration coverage for newer managed services like Agentcore and Foundry, advanced AI red-teaming and data-poisoning test cases, and permission models built specifically for MCP servers. Outside those areas, the &#8220;AI security&#8221; label is mostly repackaging.</p><h2>Three integration points, three different blind spots</h2><p>The report describes where AI security vendors plug into your environment: network proxies, endpoint agents, and hooks or gateways (like inference hooks). Each one has a real trade-off, not just a deployment preference.</p><p>A network proxy works with most agent types and gives you a flexible runtime engine, but it can&#8217;t see local permissions, the agent&#8217;s reasoning, or what the agent reads on disk. An endpoint agent sees permissions and local behavior in depth, but it does not reach SaaS-hosted agents or first-party agents running somewhere else. Hooks and gateways are the newest option. The report&#8217;s example is what it calls Anthropic&#8217;s inference hooks, together with Anthropic&#8217;s Compliance API. They give session-level visibility and can change an agent&#8217;s context window during the session, but not every agent platform supports them, and they can&#8217;t touch device settings.</p><p>The report&#8217;s conclusion is that no single integration point is correct. Full coverage needs all three, and the right mix depends on where your real risk is, not on which one is easiest to deploy first.</p><h2>The real difference is permissions and prevention</h2><p>The most useful point for engineers building a program is in the &#8220;Innovative Capabilities&#8221; section. Per the report, vendors differ most in how tightly a platform controls which tools and permissions an agent has <em>before</em> it acts, and less in how good their detection is.</p><p>The report shows this in four capability areas:</p><ul><li><p>Skill detonation. A Skill is a small application for an agent, sometimes just markdown, sometimes with scripts and packages. Detonation runs it in a sandbox at runtime and watches network egress and process execution before trusting it. The reason is that static analysis stops working once a Skill can fetch external content or change its own behavior. Dynamic analysis has its own limit, since code can try to detect the sandbox.</p></li><li><p>MCP and Skills marketplaces. Governed, self-service flows for finding and vetting agent capabilities across teams, instead of every team deciding on its own what an agent may install and call.</p></li><li><p>Granular, identity-aware permissions. Mapping an agent&#8217;s local file-system access to its cloud IAM role, so the blast radius is understood end to end.</p></li><li><p>Intent-based detection. Comparing an agent&#8217;s system prompt with its expected goals and the organization&#8217;s guardrails, then watching the live session for drift. The report adds a caveat that every evaluation team should hear: &#8220;many runtime detection engines are not as powerful as promised.&#8221; Test this hands-on before you buy it.</p></li></ul><p>All four put prevention ahead of detection. The report treats supply chain risk from MCP servers and Skills as more urgent than any exotic new AI-only threat. Many MCP servers are distributed as npm packages, so they carry the same risk as any other open-source package that could ship malware.</p><h2>What this means for how you buy</h2><p>The report&#8217;s buyer&#8217;s guide is built on one distinction that is worth adopting in your own roadmap: first-party agents (your own agents, for example built on LangChain or CrewAI) and third-party agents (Claude Code, Codex and similar tools on workforce endpoints).</p><p>For first-party agents, the advice is proportionality. A simple internal chat tool with no access to real customer data has limited risk and is probably already covered by your application and cloud security stack. Complexity changes this. Once a first-party agent has real data access or takes actions that matter, you need real AI red-teaming and runtime defense, not just posture checks.</p><p>For third-party agents, the report describes the first real decision. If AI security is a dedicated priority and you need deep control over employee endpoints, work with an endpoint specialist. If you want AI governed as part of a wider program that several teams already touch, use a broader AI security platform. Most teams end up doing both: one endpoint tool for general workforce AI use, and a specialist tool for developers. The reason is that traditional EDR and MDM tools still don&#8217;t understand IDE plugin or developer supply chain threats.</p><h2>Where this is heading</h2><p>The report predicts that the &#8220;hosted endpoint&#8221;, meaning agents that run on managed platforms like Agentcore and Foundry, will be in 2027 what the &#8220;agentic endpoint&#8221; is in 2026. So the current effort to secure agents running directly on laptops may be a transition phase. If more organizations move agent deployment to managed cloud platforms, the control point moves there too, from the devices to the platforms that host the agents.</p><p>The report&#8217;s advice for programs being built now is not to over-invest in point solutions tied to today&#8217;s endpoint-agent problem. It favors vendors and architectures that move with the frontier model providers, who ship new control surfaces like inference hooks faster than most vendors can build around them. The trade-off is that tying controls closely to one model provider&#8217;s features can make it harder to switch providers later.</p><p>A practical way to use the report is its own framework: first-party versus third-party, posture versus runtime, and proxy versus endpoint versus hook. Go through your current stack with these three questions and mark what is really missing, and what you already own under a different label.</p><div><hr></div><p>Reference: <a href="https://www.latio.com/reports/2026-latio-ai-security-report">2026 AI Security Market Report, Latio</a></p>]]></content:encoded></item><item><title><![CDATA[LLM Jailbreaking: How Users Trick AI Into Ignoring Its Own Rules]]></title><description><![CDATA[How role-play, fiction, long prompts and encodings get around safety training, why they keep working, and which defenses help]]></description><link>https://www.hackerspot.net/p/llm-jailbreaking-how-users-trick</link><guid isPermaLink="false">https://www.hackerspot.net/p/llm-jailbreaking-how-users-trick</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Tue, 29 Sep 2026 06:11:54 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!3e2w!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F11637665-eb24-40bf-b666-60721a8dca9e_923x463.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to explain how jailbreaks work. A jailbreak is a prompt that makes a chat model ignore its safety rules. The model was trained to refuse things like weapon instructions or malware, but training is not a lock, and people find ways around it. Some do it out of curiosity, some with bad intent.</p><p>Jailbreaking is different from prompt injec&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/llm-jailbreaking-how-users-trick">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[The AI Vulnerability Flood Isn’t New Bugs. It’s Old Bugs at New Scale]]></title><description><![CDATA[Stop closing AI-era vulnerability tickets one by one. Trace each finding to its root cause and fix the class.]]></description><link>https://www.hackerspot.net/p/the-ai-vulnerability-flood-isnt-new</link><guid isPermaLink="false">https://www.hackerspot.net/p/the-ai-vulnerability-flood-isnt-new</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Fri, 25 Sep 2026 01:59:08 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!QG3l!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9f11e65d-65ee-4029-be05-dc9fa4c08033_1024x559.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>We have good tools to find vulnerabilities: static analysis, fuzzing, threat modeling, SDLC gates. For a while it felt like they were closing the gap between how fast code got written and how fast it got checked. But they were built for a world where people write code at human speed. AI assistants changed that speed. In this post, I want to talk about w&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/the-ai-vulnerability-flood-isnt-new">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[What Is Jev? TypeSafe AI's "System One Model"]]></title><description><![CDATA[A fast, cheap model that returns typed decisions instead of text. What it does, how to call it, and which claims are still unverified.]]></description><link>https://www.hackerspot.net/p/what-is-jev-typesafe-ais-system-one</link><guid isPermaLink="false">https://www.hackerspot.net/p/what-is-jev-typesafe-ais-system-one</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Thu, 24 Sep 2026 00:54:17 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!l5EU!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to explain what Jev is, how to call it, how people use it next to Claude Code, and which claims about it are still unverified.</p><p>Jev launched on September 15, 2026, so most of what is written about it comes from TypeSafe itself. Every speed and cost number below is self-reported unless stated otherwise. The details were last checked on November 11, 2026.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!l5EU!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!l5EU!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg 424w, https://substackcdn.com/image/fetch/$s_!l5EU!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg 848w, https://substackcdn.com/image/fetch/$s_!l5EU!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!l5EU!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!l5EU!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg" width="1024" height="724" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:724,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;System 2 LLM versus System 1 Jev&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="System 2 LLM versus System 1 Jev" title="System 2 LLM versus System 1 Jev" srcset="https://substackcdn.com/image/fetch/$s_!l5EU!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg 424w, https://substackcdn.com/image/fetch/$s_!l5EU!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg 848w, https://substackcdn.com/image/fetch/$s_!l5EU!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!l5EU!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc8731f-e823-46cb-bb34-309df96d72cc_1024x724.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>What is Jev?</h2><p>Jev is a model from TypeSafe AI, a San Francisco startup founded in 2024. It went into limited early access on September 15, 2026, together with a 40 million dollar seed round led by DCVC. Forbes reported a valuation of around 200 million dollars, citing a person familiar with the deal, not the company. On October 9, 2026, TypeSafe announced an 870 million dollar Series A led by Andreessen Horowitz, with Sequoia Capital and DCVC, at a 7.5 billion dollar valuation.</p><p>TypeSafe calls Jev the first <strong>&#8220;System One Model&#8221;</strong>. An LLM like Claude or GPT writes text, token by token, for a person to read or for code to parse afterwards. Jev does not write text. You give it some state (a string, a JSON object or an array) and one or more typed questions, and it returns typed answers with probabilities. Software reads the answer directly.</p><p>So if your task ends in &#8220;write me a sentence&#8221;, Jev is not for you. If it ends in &#8220;is this true or false&#8221;, &#8220;which of these five categories is it&#8221; or &#8220;how confident are you&#8221;, that is the shape of task Jev targets.</p><p>The name comes from William Stanley Jevons, the 19th-century economist behind the Jevons paradox: when using a resource becomes more efficient, total use of it can go up, not down. TypeSafe&#8217;s founder uses this as the thesis for Jev. If machine decisions get much cheaper, people will build more with them, and AI spend grows.</p><p>TypeSafe was founded by Diogo Almeida (CEO), Erik Gafni and Sasha Sheng. Almeida worked at OpenAI on RLHF, InstructGPT and ChatGPT before leaving in 2024. Jev was reportedly built in stealth for about two years.</p><h2>How Jev works</h2><p>TypeSafe takes its terms from Daniel Kahneman&#8217;s <em>Thinking, Fast and Slow</em>. System 2 is slow, deliberate reasoning, and TypeSafe puts today&#8217;s chain-of-thought LLMs there. System 1 is fast, intuitive judgment, and that is where it puts Jev. This is a marketing frame. Chain-of-thought LLMs are not really System 2 thinking, but the split is useful for seeing what Jev is for.</p><p>A call has two parts. The state is the context the decision is about: a document, a conversation, a tool call log or structured data. The questions each have a fixed answer type. In TypeSafe&#8217;s API there are four: <code>noul</code> (a yes/no question that returns a probability), <code>choice</code> (one option from a list you define), <code>score</code> (a level on a rubric you define) and <code>box</code> (a location in an image, still experimental).</p><p>For example, take a support ticket. This request is from TypeSafe&#8217;s API docs:</p><pre><code><code>{
  "model": "jev-latest",
  "state": "I was charged twice for my subscription this month. Please refund the extra charge.",
  "questions": {
    "billing": {
      "type": "noul",
      "instructions": "Is this message about billing?"
    }
  }
}</code></code></pre><p>It goes to <code>POST https://api.typesafe.ai/v1/systemone</code> with an <code>Authorization: Bearer</code> header. The documented response looks like this:</p><pre><code><code>{
  "model": "jev-latest",
  "answers": {
    "billing": {
      "type": "noul",
      "noul": 0.98
    }
  },
  "usage": {
    "input_tokens": 120,
    "output_tokens": 12
  }
}</code></code></pre><p>You can add more questions to the same call, for example a <code>choice</code> for urgency and a <code>noul</code> for &#8220;does a human need to look at this&#8221;. Then your code routes the ticket on the numbers. Both examples are from the docs, not from a live run.</p><p>Per TypeSafe, all questions are evaluated against the state in parallel, in one pass, and the answers always match the schema you defined. That guarantee covers the format only: a choice is always one of your options. It says nothing about whether the answer is right. The model can still give a high probability to the wrong answer.</p><p>On training, TypeSafe has not published the architecture, the parameter count or a paper. Its launch post talks about &#8220;a new model architecture&#8221; and a &#8220;parallel sampler&#8221;. Other coverage says TypeSafe describes it as transformer-based and trained only on synthetic data. Training uses a method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD), which optimizes the output probabilities against real outcomes instead of human rater preference (the way RLHF does for chat models). Some commentators guess that Jev sits on top of an existing open-weight model. None of this is confirmed.</p><h2>How to access Jev</h2><p>There is no way to run the real Jev locally or self-host it. The weights are not published, and TypeSafe says every customer uses the same weights, with no per-customer fine-tuning.</p><p>These are the ways to call it:</p><ul><li><p>Direct API. Sign in at <code>console.typesafe.ai</code>, create a key and prepay credits. The documented endpoint is <code>POST /v1/systemone</code>. Some early sources cited <code>/api/v1/decide</code>, so if you see that path in a blog post, check <code>docs.typesafe.ai</code> first.</p></li><li><p>SDKs. <code>typesafe-sdk</code> for Python (<code>pip install typesafe-sdk</code>), which reads the key from <code>TYPESAFE_API_KEY</code>, and <code>@typesafe-ai/sdk</code> for JavaScript (<code>npm install @typesafe-ai/sdk</code>).</p></li><li><p>Cloudflare Workers AI, as <code>typesafe/jev</code>, called with <code>env.AI.run('typesafe/jev', {...})</code> and billed through your Cloudflare account. Cloudflare&#8217;s listing shows a 32k context window.</p></li><li><p>Vercel AI Gateway, as <code>typesafe-ai/jev</code>, also with a 32k context window in its listing.</p></li><li><p>Opper.ai, a third-party gateway (<code>typesafe/jev-1.14.0</code>). Sign-up is free, but Jev is a premium model there, so you need a card on file. Opper says its Jev route keeps request content in provider logs, though it is not used for training.</p></li></ul><p>At launch some sources also listed Jev on OpenRouter as <code>typesafe/jev-1.13</code>. On November 11, 2026, it was not in OpenRouter&#8217;s public model list.</p><p>Access has been reported inconsistently. Some sources say TypeSafe removed the early-access waitlist on September 20, 2026 (&#8221;available to everyone, no waitlist&#8221;). Others describe a batch-invite queue. Check TypeSafe&#8217;s own site before you plan around it.</p><p>There are also community projects like OpenJev and its successor SemIf. They copy Jev&#8217;s request and response contract (state plus typed questions in, typed answers out) using a small open model, and they run fully offline. Their own documentation is clear that they are not Jev. They copy the interface, not the model or its training. On October 1, 2026, Cloudflare also released Clef (27B) and Clef-flash (9B), its own decision models with open weights under Apache 2.0 on Hugging Face. They follow the same System One API, so if you need to self-host, they are the closest option. Cloudflare&#8217;s benchmarks for them are its own.</p><h2>Pricing</h2><p>TypeSafe&#8217;s models page lists these numbers:</p><ul><li><p>Input tokens: <code>$0.042</code> per million (that is <code>$42</code> per billion).</p></li><li><p>Output tokens: free.</p></li><li><p>Context: 64k tokens per request (state plus all questions) and 32k tokens for the state plus any one question. A <code>TypeSafe-Danger-Zone-Long-Context: true</code> header raises both to 262k, at the same price per token, but TypeSafe says it is slower and less accurate. Gateways can be stricter: Cloudflare and Vercel list 32k.</p></li></ul><p>TypeSafe charges only for input tokens. Typed answers and probabilities still come back, and the response even reports <code>output_tokens</code>, so the zero output price is a pricing decision rather than a technical consequence. TypeSafe also says in its launch post that it cannot prove the current pricing is not subsidized.</p><p>The same <code>$0.042</code> rate appears on TypeSafe&#8217;s docs, Cloudflare, Vercel and Opper. The product is new, so check the live rate before you budget anything at scale.</p><p>To get a feel for the numbers, take a decision with a 1,500-token state. On Jev the input costs:</p><pre><code>1,500 x $0.042 / 1,000,000 = $0.000063</code></pre><p>A million of those decisions costs <code>$63</code>. The same 1,500 tokens sent to an LLM at <code>$3</code> per million input tokens cost <code>$0.0045</code>, about 71 times more, and that is only the input. The LLM also charges for every output token. Prices change, so plug in your own.</p><p>TypeSafe&#8217;s cost tips are practical. Trim the state to what the decision needs, since you pay per input token. Ask several questions against one state in a single call instead of making repeated calls. And route by confidence: let Jev settle the high-confidence cases and send only the low-confidence ones to a full LLM or a human.</p><h2>Speed, cost and fit</h2><p>TypeSafe says a call takes 70 to 500 ms end to end. For a decision that takes a frontier LLM a few seconds, that would be a big difference. Its launch post claims Jev is 40x to 200x faster than frontier models at the same quality on &#8220;System One-shaped&#8221; queries. The headline &#8220;193.6x faster&#8221; and &#8220;444.6x cheaper&#8221; come from its own workflow evals, which the post itself calls &#8220;on the higher end of real world gains&#8221;. The same post lists its own caveats: the speed tests ran from the team&#8217;s laptops, the team built the workflows, and the reference answers come from other vendors&#8217; models.</p><p>The schema guarantee also means there is no free text to parse and no malformed JSON to retry. TypeSafe&#8217;s &#8220;0% type errors&#8221; is true by construction, not a measured result.</p><p>The best fit is as a pre-filter. Jev does not replace your LLM. It sits in front of it and does the cheap triage, classification and routing, so the expensive model only gets the requests that need generation or reasoning.</p><p>On the other side, long documents need chunking or a retrieval step in front of Jev. The weights are closed, so you only know what TypeSafe tells you about architecture, size and training data. And everyone calls the same model.</p><h2>Using Jev with Claude Code</h2><p>Since Jev does not write text, it does not replace Claude. People use it as a cheap decision step inside a harness that still calls Claude for real generation or reasoning. Two shapes have come up.</p><p>The first is routing. Jev looks at the request first. Does it need a full model call at all? Which sub-task is it? How confident are we already? Only then, if needed, Claude does the work. In an agent loop this keeps frontier prices for the turns that need a frontier model.</p><p>The second is context compaction. A community plugin, <code>fast-jev-compaction</code>, replaces Claude Code&#8217;s built-in <code>/compact</code>. The built-in version asks the model to write a lossy summary of the old conversation. The plugin instead asks Jev to score every tool call and result outside a protected recent window, and drops or truncates the ones below a keep-threshold. Whatever it keeps stays verbatim, never rewritten. That fails differently from a summary that silently drops a file path or an error string. The plugin&#8217;s author also warns that a probability is not proof that a result is safe to delete.</p><p>Before you install this plugin, read these points.</p><p>It is a third-party plugin, not built or endorsed by Anthropic or TypeSafe. It runs inside Claude Code with your permissions, and it sends your conversation, including tool calls and their output, to TypeSafe&#8217;s API. That output can contain code, file contents and secrets. Read the source first, install from the original repository and not one of its copies, and check TypeSafe&#8217;s data retention terms.</p><p>It was built on Claude Code &#8220;function hooks&#8221;. These come from an Anthropic proposal, GitHub issue <code>anthropics/claude-code#91870</code>, opened on September 3, 2026. When this post was first written, the feature was a prototype behind the default-off flag <code>CLAUDE_CODE_ENABLE_FUNCTION_HOOKS</code>, and details like event names and hook failure behavior were still open. It has since shipped. The official changelog for Claude Code 2.1.287 says &#8220;Added Claude Mods: plugins may now modify deeper behavior&#8221;. It does not say whether the flag is still needed, so if the plugin shows no hooks on a current build, try setting it.</p><p>My first version of this post said there was an open bug report, <code>tamaratran/fast-jev-compaction#76</code>, about the hooks not loading on Claude Code 2.1.278. That issue was opened and closed by the reporter on September 21, 2026, with no explanation and no fix linked. If the plugin shows &#8220;Hooks (0)&#8221; for you, check your Claude Code version first.</p><p>The plugin README asks for Claude Code 2.1.274 or later and the flag in <code>~/.claude/settings.json</code>, like this:</p><pre><code><code>{
  "env": {
    "CLAUDE_CODE_ENABLE_FUNCTION_HOOKS": "1"
  }
}</code></code></pre><p>The README also puts <code>TYPESAFE_API_KEY</code> in that same <code>env</code> block. Don&#8217;t do that. The plugin reads the key from the environment by default, so set it in your shell or secret manager instead of a settings file that might get synced or committed:</p><pre><code><code>export TYPESAFE_API_KEY="&lt;your key&gt;"</code></code></pre><p>Then install it:</p><pre><code><code>claude plugin marketplace add tamaratran/fast-jev-compaction
claude plugin install fast-jev-compaction@fast-jev-compaction</code></code></pre><p>Restart Claude Code or run <code>/reload-plugins</code> afterwards. The documented defaults are <code>keepThreshold</code> 0.5, <code>preserveRecentMessages</code> 6, <code>maxStateTokens</code> 25,000, <code>maxRequestTokens</code> 30,000 and <code>truncateHeadChars</code> 300.</p><h2>A warning about the coverage</h2><p>Many sites write about Jev, and a lot of them repeat the same numbers word for word: &#8220;193.6x faster&#8221;, &#8220;444.6x cheaper&#8221;, the same GitHub star counts and Hacker News vote counts. Most of these are new sites that appeared within days of each other. That pattern is typical of content-farm copy, not independent reporting. The speed and cost numbers start with TypeSafe; the farms only repeat them. Jev itself is real. Forbes, TechCrunch and The Register covered the launch, and TypeSafe&#8217;s docs are public. But none of the performance claims has been reproduced in a published third-party benchmark as of this writing, and the community tooling is unofficial and early.</p><p>If you try it, use a separate API key with a spend cap, check the live price and limits on <code>docs.typesafe.ai</code>, and read the plugin source before you install anything.</p>]]></content:encoded></item><item><title><![CDATA[Model Extraction: How Attackers Steal Your AI Through Its API]]></title><description><![CDATA[A model extraction attack lets attackers copy your AI through its own API. How surrogate models are built from queries, and which defenses raise the cost.]]></description><link>https://www.hackerspot.net/p/model-extraction-how-attackers-steal</link><guid isPermaLink="false">https://www.hackerspot.net/p/model-extraction-how-attackers-steal</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Mon, 21 Sep 2026 16:07:44 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!yAdR!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1e9d47a3-aa9c-4201-8dba-1f01cf3cab7c_1024x687.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to explain model extraction. If your model sits behind an API, anyone who can send queries can record the answers and train a copy of it. The attacker&#8217;s budget is the price of your API calls, which is usually far less than your training bill.</p><p>This post is about prediction APIs, such as classifiers and scoring models. For a small mode&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/model-extraction-how-attackers-steal">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Backdoor Attacks: The Hidden Trigger Inside Your AI Model]]></title><description><![CDATA[A backdoored model passes your tests and only misbehaves when it sees a secret trigger. How it works, and what you can check]]></description><link>https://www.hackerspot.net/p/backdoor-attacks-the-hidden-trigger</link><guid isPermaLink="false">https://www.hackerspot.net/p/backdoor-attacks-the-hidden-trigger</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Mon, 14 Sep 2026 17:25:57 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!MDop!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to explain backdoor attacks on AI models. The idea is simple. The model works normally, until the input has a secret trigger. Then it gives the answer the attacker wants.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!MDop!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!MDop!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!MDop!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!MDop!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!MDop!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!MDop!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg" width="1024" height="687" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:687,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!MDop!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!MDop!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!MDop!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!MDop!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b0bcac-f105-4d36-b6f9-19608af5a4bb_1024x687.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>It is like a lock that was built with one extra key, except the model learns its extra key during training. Researchers have shown this many times, and it matters f&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/backdoor-attacks-the-hidden-trigger">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Data Poisoning: How Attackers Corrupt AI Before It Even Launches]]></title><description><![CDATA[How label flipping, input tampering and backdoor triggers work, what real research shows, and the controls that help]]></description><link>https://www.hackerspot.net/p/data-poisoning-how-attackers-corrupt</link><guid isPermaLink="false">https://www.hackerspot.net/p/data-poisoning-how-attackers-corrupt</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Mon, 07 Sep 2026 17:22:53 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!whuM!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>To poison an AI model, the attacker does not need your servers or your code. They only need to change some of the data you train on. In this post, I want to explain how that works, what research has shown so far, and what you can do about it.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!whuM!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!whuM!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!whuM!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!whuM!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!whuM!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!whuM!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg" width="1024" height="687" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:687,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!whuM!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!whuM!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!whuM!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!whuM!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe336cd2e-dc3c-43eb-b535-1867055a43b2_1024x687.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>Data poisoning means adding malicious or corrupted samples to a training set so the model learns the wrong behavi&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/data-poisoning-how-attackers-corrupt">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Your AI Has a Supply Chain and It Can Be Compromised]]></title><description><![CDATA[Models and data are a second supply chain. How attackers use it, and what to pin and check]]></description><link>https://www.hackerspot.net/p/your-ai-has-a-supply-chain-and-it</link><guid isPermaLink="false">https://www.hackerspot.net/p/your-ai-has-a-supply-chain-and-it</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Mon, 31 Aug 2026 17:20:09 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!5jvJ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to look at what your AI system depends on. It runs on frameworks like PyTorch or TensorFlow. It uses pre-trained models from the internet and datasets from many places. Any of them can be compromised, and most teams don&#8217;t check them.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!5jvJ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!5jvJ!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp 424w, https://substackcdn.com/image/fetch/$s_!5jvJ!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp 848w, https://substackcdn.com/image/fetch/$s_!5jvJ!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp 1272w, https://substackcdn.com/image/fetch/$s_!5jvJ!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!5jvJ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp" width="1024" height="687" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/c3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:687,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!5jvJ!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp 424w, https://substackcdn.com/image/fetch/$s_!5jvJ!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp 848w, https://substackcdn.com/image/fetch/$s_!5jvJ!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp 1272w, https://substackcdn.com/image/fetch/$s_!5jvJ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3f1320d-08ee-4159-9d33-329101faf6ca_1024x687.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>AI has two supply chains. One is code and libraries. SolarWinds showed in 2020 how bad this one can g&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/your-ai-has-a-supply-chain-and-it">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[The AI Hype Cycle: How to Tell Real Progress from Marketing Noise]]></title><description><![CDATA[AI has been overhyped and defunded before. Four questions that help separate real progress from marketing, and what the history teaches.]]></description><link>https://www.hackerspot.net/p/the-ai-hype-cycle-how-to-tell-real</link><guid isPermaLink="false">https://www.hackerspot.net/p/the-ai-hype-cycle-how-to-tell-real</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Mon, 24 Aug 2026 17:20:14 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!1A9C!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to look at the AI hype from a security point of view. We have seen this before. In the 1950s and 60s, AI researchers promised machines with human-level intelligence within a generation. By the mid-1970s the funding was gone, and the field went into what is now called the first &#8220;AI winter&#8221;.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!1A9C!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!1A9C!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!1A9C!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!1A9C!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!1A9C!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!1A9C!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg" width="1024" height="687" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/b7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:687,&quot;width&quot;:1024,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!1A9C!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg 424w, https://substackcdn.com/image/fetch/$s_!1A9C!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg 848w, https://substackcdn.com/image/fetch/$s_!1A9C!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!1A9C!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7bb0363-c145-4ac5-aee5-31865fac3326_1024x687.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>It happened again in the 1980s with expert s&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/the-ai-hype-cycle-how-to-tell-real">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[When to Use AI - A Simple Litmus Test Before You Build]]></title><description><![CDATA[Four questions to ask before you add a model to the stack]]></description><link>https://www.hackerspot.net/p/when-to-use-ai-a-simple-litmus-test</link><guid isPermaLink="false">https://www.hackerspot.net/p/when-to-use-ai-a-simple-litmus-test</guid><dc:creator><![CDATA[Chady]]></dc:creator><pubDate>Mon, 17 Aug 2026 15:55:41 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!wNyg!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to share four questions to ask before adding AI to a project. Many problems do not need it. A model brings new dependencies and new ways to fail, so it should earn its place. Once you count data, infrastructure, monitoring and maintenance, an AI solution can cost much more than the rule or query it replaced.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!wNyg!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!wNyg!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg 424w, https://substackcdn.com/image/fetch/$s_!wNyg!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg 848w, https://substackcdn.com/image/fetch/$s_!wNyg!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!wNyg!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!wNyg!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg" width="964" height="538" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:538,&quot;width&quot;:964,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!wNyg!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg 424w, https://substackcdn.com/image/fetch/$s_!wNyg!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg 848w, https://substackcdn.com/image/fetch/$s_!wNyg!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!wNyg!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff6be5522-e943-4ebe-b827-11dfafa08104_964x538.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>The four questions</h2><p>Run the&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/when-to-use-ai-a-simple-litmus-test">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Hack the Planet: Walking the CyberDelias Gauntlet]]></title><description><![CDATA[A DEF CON Cloud Village CTF write-up, from an AI chatbot through Log4Shell and an AWS PassRole chain to a phone call]]></description><link>https://www.hackerspot.net/p/hack-the-planet-walking-the-cyberdelias</link><guid isPermaLink="false">https://www.hackerspot.net/p/hack-the-planet-walking-the-cyberdelias</guid><dc:creator><![CDATA[Hackerspot Team]]></dc:creator><pubDate>Sat, 15 Aug 2026 00:13:18 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!wvRL!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F75237c55-d908-4e6d-9693-5fa49449b136_1024x633.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In this post, I want to share my walkthrough of the CyberDelias challenge. It has an AI chatbot, Log4Shell, an AWS privilege escalation and, at the end, one phone call.</p><p>The challenge is based on the movie <em>Hackers</em> (1995), where a crew of teenage hackers shouts &#8220;Hack the planet!&#8221; and hangs out at a neon club. Here the club is CyberDelias, &#8220;The Geek &amp; Hacke&#8230;</p>
      <p>
          <a href="https://www.hackerspot.net/p/hack-the-planet-walking-the-cyberdelias">
              Read more
          </a>
      </p>
   ]]></content:encoded></item></channel></rss>