Cato’s AgentSense Targets AI Agents Before They Go Rogue

AI agents can take harmful actions. Cato Networks has unveiled AgentSense, a runtime security capability designed to stop those actions before they execute. Published November 17, 2026, the announcement positions the system as a defense against indirect prompt injection — the attack class that turns an agent’s own information sources against it.
Cato AI Labs developed AgentSense to trace how untrusted content influences an agent’s behavior. Cato says the capability is integrated into its platform and will be included at no additional charge within Cato AI Usage licenses.
That matters because valid credentials do not guarantee safe behavior. An agent can hold the right permissions and still be manipulated into harmful actions, turning a routine workflow into an unauthorized detour with access already built in.
The Permission Problem Is Only Half the Problem
An enterprise agent may need to read emails, search documents, consult websites, and call tools to complete a task. Every source expands the information available to the agent, but that same boundary creates room for indirect prompt injection.
A document, ticket, web page, or tool response can contain text that attempts to redirect the workflow, expose information, or trigger an unrelated action. The content does not need to arrive as a direct instruction from the user; it can enter through the material the agent was asked to process.
Checking whether an agent can access an application remains important, but it answers a different question from whether a particular action serves the task the user authorized. An agent may have permission to reach a system while lacking a valid reason to perform a specific sensitive operation.
AgentSense’s proposition is to evaluate that behavior during execution, with an opportunity to intervene before a sensitive action is carried out. Cato says the system blocks an action when it detects untrusted material steering the agent toward an unsafe step.
Tracing Influence Instead of Matching Known Attacks
The announced Untrusted Influence Detection capability separates trusted instructions from untrusted content within an AI session and traces how the untrusted material affects behavior and outputs. The focus is not only what the agent can see, but how that material changes the path it follows.
Cato describes AgentSense as assessing influence rather than relying on static classifiers or a catalog of previously observed attack patterns. That distinction gives the system a broader target: the relationship between untrusted content and an agent’s next action, not just the wording of a known attack.
The approach also reflects the awkward reality of agent security. A workflow can begin with a legitimate user request, use legitimate tools, and still end with a harmful action after untrusted content redirects the process. Security teams therefore need to examine execution itself, not only permissions or input categories.
Cato describes the technology as patented, but it has not provided a patent identifier. The company has paired the launch with an invitation to test the claim rather than asking everyone to accept another security announcement on faith.
The adversarial testing challenge offers $10,000 each to the first three eligible participants who successfully redirect a protected workflow into malicious actions. The contest specifically tests whether AgentSense can resist workflow redirection, placing the system’s central promise in the hands of researchers trying to break it.
That challenge will not settle every question about runtime protection, but it creates a clear test of the advertised defense. If AgentSense can detect untrusted influence before a protected workflow performs a malicious action, Cato will have addressed a weakness that application access controls alone cannot answer.
Based on
- Cato Networks Unveils AgentSense to Stop Harmful AI Agent Actions Before Execution — unite.ai
- BeyondTrust Experts Reveal Top Cybersecurity Predictions for 2027 and Beyond | Markets Insider — markets.businessinsider.com
- Incode Launches Incode Know Your Agent to Anchor AI Agent Trust in Verified Human Identity | AFP — afp.com
- Menlo Security IRIS Browser Extension Lets Security Teams Rein In Shadow AI and Risky Extensions Without a Browser Swap | AFP.com — afp.com




