Felix ReportIndependent Reporting and News

Cato opens an AI-agent challenge with a narrow test of its protection claims

Conceptual illustration of blue agent symbols inside a transparent boundary with documents approaching it.

The Israeli-founded security company is offering three $10,000 prizes for qualifying sandbox results. The contest tests a defined workflow, not a promise that every production agent is protected.

Cato Networks opened its AgentSense challenge on November 17 at 3:30 p.m. Israel time, inviting participants to test whether an AI agent can be redirected through material it reads. The launch accompanies the company's announcement of technology intended to stop harmful agent actions before execution.

The useful distinction is between the claim and the experiment. Cato describes a system that traces the influence of untrusted content on an agent's behavior. The challenge gives eligible participants a bounded environment in which to test that proposition. It does not establish an independent success rate across customers' systems.

A test inside a defined environment

The >official challenge page (opens in a new tab) requires completing both levels, a chatbot copilot and a coding agent, and requires qualifying attempts to succeed with AgentSense enabled. Its terms describe fictional challenge data and restrict testing to the assigned sandbox. They do not authorize testing Cato's production infrastructure or third-party systems.

The first three eligible successful participants can receive $10,000 each, subject to Cato's validation. The deadline is December 8 at 7 a.m. Israel time. Participation has age, country and other eligibility conditions; an apparent technical success alone is not a confirmed prize award.

Permission to act is not the same as the right instruction

In its product announcement, Cato argues that an agent can hold valid privileges yet take an action outside the task a user intended. An email or document can contain information the agent needs while also carrying instructions that should not control it. The company says AgentSense separates these roles and blocks unsafe actions associated with untrusted influence.

That is a company description of its approach. It should not be read as evidence that every hidden instruction is detected or that legitimate work will never be interrupted. Those are questions for testing in the relevant workflow, including how the system handles ordinary documents and allowed actions.

What the announcement leaves to evaluation

Cato also describes a drift-detection capability in future tense, intended to identify actions that depart from an original task. The distinction matters when comparing what the company says its current protection does with what it says an additional capability will do. The announcement is not a license to treat the entire roadmap as already deployed.

For organizations considering agents, the launch offers a concrete evaluation question: can a permitted action be connected to the task that justified it? The challenge may generate useful examples, but its results will remain tied to the specified environment and judging rules. Buyers would still need to examine deployment scope, operational behavior and evidence from their own authorized tests before drawing broader conclusions.