Cato Networks Adds Real-Time AI Agent Controls, Says Zero Trust Falls Short
Cato's AgentSense traces what shapes an AI agent's decisions at runtime and blocks actions that drift from its mandate. Here's what SASE teams should weigh.



Image: Anton Grabolle / Better Images of AI / CC BY 4.0
Cato Networks is arguing that zero trust, on its own, no longer covers the risk that AI agents introduce. Its answer is to push monitoring into the secure access service edge (SASE) platform so that an agent's behaviour is assessed while it runs, not only when it connects or authenticates. For network and security teams already rolling out copilots and autonomous tools, that's a meaningful shift in where the control point sits.

What AgentSense actually does
The new capability is called Cato AgentSense. According to Cato, it traces how instructions and data feed into an AI agent's behaviour at runtime. It also flags when the agent starts to move away from the task it was originally given. When a deviation appears, the service blocks the deviant action. The stated target is indirect prompt injection, where an agent is steered by content it reads from an outside source rather than by the instructions its operator wrote.
That distinction matters more than it might first appear. Traditional access controls answer the question "who is this user, and can they reach this resource?" An agent that summarises email, reads documents or calls external tools can be pulled into an unwanted action by text it never should have treated as a command, even when its identity and network path are perfectly legitimate. Checking the connection tells you nothing about the instructions the agent is acting on.

Itay Ravia, head of Cato AI Labs, framed the goal in terms of trust in the outcome: "These examples underscore why enterprises need to trust that an AI agent will carry out the task they authorized, even when it encounters content an attacker controls." He added that AgentSense "helps preserve that trust by tracing how untrusted content influences an agent's decisions and sensitive actions, helping enterprises adopt AI agents without giving attackers control of the workflow."
The attack examples Cato is pointing to
Cato cited three recent cases to make the case that the threat is concrete. The first is EchoLeak, which the company describes as a zero-click vulnerability in Microsoft 365 Copilot. According to Cato, it allowed attackers to quietly steal sensitive organisational data without the victim taking any action. A zero-click flaw is the kind that worries administrators most, because there's no user mistake to train away.

The second is CurXecute, a remote code execution flaw disclosed in 2025 that targets the Cursor AI code editor through its model context protocol (MCP) integration. MCP is quickly becoming the plumbing that lets agents connect to tools and data sources, which means every MCP connection is also a path that untrusted content could reach.
The third is DuneSlide, a set of flaws that Cato says let prompt-injected instructions get around Cursor's built-in sandbox protections and trust boundaries. The lesson Cato draws is that an agent's own guardrails can't be the only layer. When the sandbox itself can be talked around, something outside the agent has to observe what it's actually doing.

Where it fits and what it costs
AgentSense is integrated into Cato's flagship SASE platform. Cato says it is included at no additional charge within Cato AI Usage licences. If you're already on a Cato AI Usage licence, that removes the procurement question for the agent-monitoring feature itself. It doesn't, however, answer whether the capability meets your requirements for coverage, logging or response, and those are the questions worth taking into a proof of concept.
The service has roots in Cato's $320 million purchase of Aim Security, a firm focused on securing enterprise use of large language models (LLMs) and generative AI. That lineage explains the emphasis on prompts, responses and agent actions rather than on packets and flows, which is where a traditional SASE product has historically looked.

Shlomo Kramer, Cato's CEO, has argued that today's security stacks need to "cover the prompts and responses and action of agents across the muscle of the organization." He described a full life cycle running from posture management through protection and detection response, then back to posture. He went further, saying, "This is a completely new security stack that is very different than the network traffic and the application and data security stacks." That's a bold claim, and it's one to test rather than accept. Many organisations will be able to fold agent controls into existing SASE, SSE and identity tooling, and the boundary between those categories is still blurry.
What I'd check before the next change window
My read is that the most useful thing this announcement does is force a question many teams haven't answered yet: which agents are running in your estate, what can they reach, and who approved that access? Here's the thing: you can't monitor what you haven't inventoried. Start there, whichever vendor you end up using.
Next, map the untrusted inputs each agent reads. Email, shared documents, web pages, tickets and third-party tool outputs all count. Any of them can carry instructions that an agent might follow. Look at where sensitive actions, such as sending data externally, modifying records or executing code, can be triggered from those inputs. Those are the paths to review first.
Then consider where enforcement would sit. Cato's approach puts it in the network and SASE layer, which suits organisations that already route agent traffic through a managed edge. Others will find more value closer to the agent or the application. Whatever you choose, test it against the scenarios Cato cites, such as zero-click data exposure in a productivity copilot and sandbox bypass in a code editor, rather than against a vendor demo.
Cato hasn't published availability dates beyond the product's current integration, and the source doesn't give independent test results for AgentSense's detection accuracy or its false-positive rate. Those numbers matter a lot when a control starts blocking actions in production. Ask for them, and ask how blocked actions are logged and surfaced to the people who will have to investigate them.
The direction of travel is clear enough. Zero trust still governs who and what gets access. The open question for the next year is how to govern what an agent decides to do once it has that access, and Cato has put a clear marker down on that problem.