Skip to content

Cato AgentSense Extends the SASE Platform to Runtime AI Agent Security

by Chris DePuy, 650 Group / November 17, 2026.

This morning, Cato Networks announced Cato AgentSense, an AI agent security service that runs inside the Cato platform. The service assesses agent behavior in real time and stops harmful actions before execution, according to Cato’s announcement. Cato says the technology is patented and was built by Cato AI Labs. In fact, Cato says it serves thousands of organizations. We counted more than 4,800 customers in our August coverage of the company’s agentic AI security push. For that installed base, Cato AgentSense arrives as a service to enable, not a product to install.

Cato AgentSense traces an untrusted README instruction through a coding agent to a stopped command
Cato AgentSense blocks an untrusted instruction before the agent executes it Source Cato Networks AgentSense product page

Cato describes it as a single toggle. The company includes it in Cato AI Usage licenses at no added cost, per the announcement. The company is putting runtime agent control inside the SASE platform enterprises already run, rather than asking them to add another point product. So that placement turns the network platform into the enforcement point for AI agent risk.

What is Cato AgentSense?

Cato AgentSense is the company’s runtime assessment layer for AI agents. It keeps agent decisions, actions, and outputs trustworthy while agents operate autonomously. Cato’s rationale for introducing AgentSense is that, while Zero Trust primarily secures access, AgentSense goes further and verifies how agents act. With Zero Trust alone, agents can hold valid privileges and still act outside the task they were given. Cato AgentSense assesses actions before execution so that drift does not become harm.

In the announcement, Cato CEO Shlomo Kramer called agent behavior control “as essential as financial controls.” Nithin Reddy, vice president of cybersecurity at Dayforce, an HR, payroll and finance software company, tied the product to a customer need. He cited the need to “detect and stop actions that could put our customers, data, or business at risk.” Organizations on Cato SASE with AI Usage licenses can enable Cato AgentSense with no new hardware or software. Enterprises standardized on other networks face a platform decision rather than a feature decision.

How does AgentSense Work?

Cato AgentSense Untrusted Influence Detection traces which instructions and data steer an agent’s behavior while it runs. First, the capability splits trusted instructions from untrusted content within the AI session. It then traces how untrusted content drives agent behavior and outputs. When untrusted content steers an agent toward an unsafe action, AgentSense then blocks it before execution.

The detection process also looks for indirect prompt injection opportunities, where attackers hide instructions in emails, tickets, documents, webpages, or tool responses. Because the tracing does not depend on known attack patterns, Cato says zero-day techniques stay in scope. Cato explains the multiple exposures as a “lethal trifecta”: sensitive data access, untrusted content, and tools that execute actions.

Itay Ravia, head of Cato AI Labs, named DuneSlide, EchoLeak, and CurXecute as still-active prompt injection cases in the announcement. He said AgentSense preserves trust “by tracing how untrusted content influences an agent’s decisions and sensitive actions.” We covered Cato’s DuneSlide disclosure in August: two Cursor IDE flaws, each scored 9.8 on the CVSS scale. Per Cato’s product page, that team disclosed EchoLeak, which let Copilot leak data through hidden instructions without any user action.

AgentSense Drift Detection

Cato AgentSense Drift Detection watches a different failure mode. As the workflow advances, it assesses whether agent actions still serve the original task or intent. Cato says it will flag individually permitted steps that fall outside the assigned task. Untrusted Influence Detection watches what drives an action; Drift Detection watches whether the action still serves the task.

The Cato AgentSense Challenge

Cato paired the launch with the Cato AgentSense Challenge, a public adversarial test. The security community is invited to attack an AgentSense-protected workflow and steer it from user intent with planted instructions. The terms run the Challenge from November 17 through December 8, 2026. Cato will pay $10,000 to each of the first three eligible participants who complete it, with no purchase required. The sandbox holds two levels: one a chatbot copilot, the other a coding agent. Moreover, a qualifying attack must arrive through indirect prompt injection and must succeed with AgentSense protection enabled. Afterward, Cato says it will use participants’ findings to strengthen the capability. We think a public test is the right pressure on a pre-execution control, and we will watch the results.

Availability

Cato AgentSense lives inside the Cato platform and needs no separate hardware or software to run, per the announcement. In other words, agent security becomes platform configuration, not a new project.

Overall, runtime agent control is converging into the network platforms enterprises already run. Cato enforces from its SASE cloud. Cato is not alone in its product expansion to include AI detection and support in its security products: the agents-and-AI theme at RSA Conference 2026 pointed the same direction earlier this year. We are eager to learn more about Cato’s customer experiences using AgentSense in the future, and for that matter, the results of the AgentSense Challenge which ends December 8. 650 Group researches several of the technologies we covered in this blog, including Enterprise, Security and SASE, and more.

author avatar
Chris DePuy
Technology Analyst and co-founder of 650 Group