Latio’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.
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.
Security moved from the browser to the endpoint
A year ago, “AI security” 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.
In 2026 the focus moved. The report’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’t give much detail on who answered, so treat these as one survey’s numbers, not an industry fact. A year earlier, endpoint agents were not even a category vendors built for. Now, the report says, “a platform is incomplete without one.”
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’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.
Budget is no longer the bottleneck
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.
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’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.
Why the vendor market is so hard to read
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 “AI red-teaming”. Identity and non-human-identity vendors describe agent permission sprawl as an AI problem. Per the report, most “agentic identity” issues are really agents using existing overpermissioned access or tokens.
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’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.
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 “AI security” label is mostly repackaging.
Three integration points, three different blind spots
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.
A network proxy works with most agent types and gives you a flexible runtime engine, but it can’t see local permissions, the agent’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’s example is what it calls Anthropic’s inference hooks, together with Anthropic’s Compliance API. They give session-level visibility and can change an agent’s context window during the session, but not every agent platform supports them, and they can’t touch device settings.
The report’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.
The real difference is permissions and prevention
The most useful point for engineers building a program is in the “Innovative Capabilities” section. Per the report, vendors differ most in how tightly a platform controls which tools and permissions an agent has before it acts, and less in how good their detection is.
The report shows this in four capability areas:
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.
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.
Granular, identity-aware permissions. Mapping an agent’s local file-system access to its cloud IAM role, so the blast radius is understood end to end.
Intent-based detection. Comparing an agent’s system prompt with its expected goals and the organization’s guardrails, then watching the live session for drift. The report adds a caveat that every evaluation team should hear: “many runtime detection engines are not as powerful as promised.” Test this hands-on before you buy it.
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.
What this means for how you buy
The report’s buyer’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).
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.
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’t understand IDE plugin or developer supply chain threats.
Where this is heading
The report predicts that the “hosted endpoint”, meaning agents that run on managed platforms like Agentcore and Foundry, will be in 2027 what the “agentic endpoint” 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.
The report’s advice for programs being built now is not to over-invest in point solutions tied to today’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’s features can make it harder to switch providers later.
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.
Reference: 2026 AI Security Market Report, Latio


