Cato Unveils AgentSense to stop Harmful AI Agent Behavior in Real-Time

Cato Networks announced Cato AgentSense, a patented breakthrough from Cato AI Labs that continuously assesses AI agents in real time to protect the integrity of their decisions, actions, and outputs as they operate autonomously.

“As enterprises entrust AI agents with decisions that shape revenue, customer relationships, and business continuity, control over agent behavior has become as essential as financial controls,” said Shlomo Kramer, CEO and co-founder of Cato Networks. “Giving agents greater autonomy requires security that can intervene before their actions cause harm. Our vision for AgentSense is to give enterprises the confidence to put agents in charge of critical business processes while retaining control over how they operate.”

Zero Trust is not enough in the AI era. Beyond securing user access, enterprises need runtime verification of how agents use their privileges. Cato delivers the converged network and security cloud for end-to-end visibility into human and AI interactions, bringing shared context directly to the point of enforcement for real-time control. Agents can have valid privileges yet still take unsafe actions outside their authorized task. AgentSense is designed to stop unsafe actions before an agent acts. It assesses actions at runtime to prevent drift from becoming harmful action.

AgentSense Untrusted Influence Detection traces how instructions and data influence an agent’s behavior at runtime. When it detects that untrusted content is steering the agent toward an unsafe action, it blocks that action before the agent executes it. This provides protection against indirect prompt injection in AI agents, allowing organizations to detect and stop attacks even when they do not match previously observed patterns. Instead of relying on static classifiers or known attack patterns, AgentSense separates trusted instructions from untrusted content inside the AI session and traces how untrusted content influences both the agent’s behavior and its outputs, providing a one-of-its-kind layer of protection from zero-day vulnerabilities.

AgentSense Drift Detection will identify when agent actions depart from the original task or intent. As a workflow progresses, it will evaluate whether the agent’s actions still serve the intended objective, helping preserve task integrity by identifying steps that may be individually permitted but fall outside the task it was given. It will complement Untrusted Influence Detection by assessing whether an agent’s actions remain aligned with its purpose, alongside understanding what influences those actions.

“DuneSlide, EchoLeak, and CurXecute are some notable examples of prompt injection vulnerabilities that continue to make headlines. 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,” said Itay Ravia, head of Cato AI Labs. “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.”

Protecting the integrity of agentic workflows
AI agents continue to scale faster than the guardrails put in place to secure them. According to Gartner, 71% of organizations expect to deploy agentic AI within the next 12 months, yet only 21% currently have agentic AI governance capabilities in place. This presents a significant security challenge, with forecasts reporting that by 2028, ungoverned AI agent abuse will drive 25 percent of enterprise breaches.

Underlying this growing risk is a fundamental architectural weakness: agents cannot reliably distinguish between trusted instructions and untrusted data they pick up as they work. Attackers can exploit this ambiguity by hiding malicious instructions in emails, webpages, documents, tickets, or tool responses, effectively redirecting agents from the user’s intended task. When those agents also have access to sensitive enterprise data and tools that can execute actions, the “lethal trifecta” turns prompt injection into a business risk with the potential for data exposure, system compromise, unauthorized actions, and business process manipulation.

Putting AgentSense to the test
Cato also introduced the AgentSense Challenge which brings the risk of prompt injection attacks to life by inviting the security community to attack an AI agent workflow protected by AgentSense and plant instructions, with the goal of steering it away from the user’s intent and into performing malicious actions. The first three eligible participants to successfully complete the Challenge will each receive $10,000 USD.

The Challenge also creates an opportunity to improve AgentSense through adversarial testing. Cato will use participants’ findings and feedback to evaluate and strengthen the capability.

Availability
Cato AgentSense is integrated within the Cato platform and will be included at no additional charge within Cato AI Usage licenses.