Cato Networks Launches AgentSense to Block AI Agent Attacks
Cato Networks has launched AgentSense, a runtime security capability that blocks harmful AI-agent actions before execution, paired with a $10,000 challenge to researchers who can break it.

Why it matters
- Cato Networks unveiled AgentSense, a runtime security capability that stops harmful AI-agent actions before they execute.
- Cato will pay $10,000 each to the first three eligible participants who successfully redirect a protected workflow into malicious actions.
- AgentSense targets indirect prompt injection by tracing how untrusted content steers an agent's actions within an AI session.
- The capability is included at no additional charge within Cato AI Usage licenses.
- Cato's CurXecute research found prompt injection could make a coding agent modify MCP configuration; Cursor fixed the flaw in version 1.3.
Cato Networks has launched AgentSense, a runtime security capability that stops harmful AI-agent actions before they execute — and it will pay $10,000 each to the first three researchers who can break it.
The company's AgentSense Challenge invites the security community to plant instructions inside a protected workflow and steer the agent away from the user's intent. Successful eligible participants take home the bounty. Cato says the findings and participant feedback will inform further evaluation and improvement of the product.
The stakes are real. Enterprises are handing AI agents legitimate credentials — email access, document search, tool calls — and those credentials do not guarantee safe behavior. An agent with valid permissions can still be manipulated into harmful actions. That gap is exactly what Cato's new technology, developed by Cato AI Labs, is built to close.
What does AgentSense actually stop?
The target is indirect prompt injection, one of the most stubborn problems in agent security.
Instead of issuing an instruction directly as the authorized user, an attacker hides one in material the agent encounters while working. A document, a ticket, a web page, or a tool response can contain text that attempts to redirect the workflow, expose information, or trigger an unrelated action.
Consider the difference between two questions. Checking whether an agent can access an application answers a permission question. It does not answer whether a particular action serves the task the user actually authorized. An agent may be permitted to read business records without being authorized to send them to a destination suggested by a retrieved document.
AgentSense's proposition is to evaluate behavior during execution, with a chance to intervene before a sensitive action is carried out. Cato says the capability is integrated into its platform and included at no additional charge within Cato AI Usage licenses.
How does untrusted influence detection work?
The core mechanism is a capability Cato calls Untrusted Influence Detection. It separates trusted instructions from untrusted content within an AI session, then traces how the untrusted material affects the agent's behavior and outputs.
Cato says AgentSense blocks an action when it detects untrusted content steering the agent toward an unsafe step.
The distinction matters because an attack does not have to look like a familiar malicious prompt. Cato describes AgentSense as assessing influence rather than relying on static classifiers or a catalog of previously observed attack patterns. That design is intended to catch new forms of indirect prompt injection, not just recognizable ones.
The release describes the technology as patented, but it provides no patent identifier, no detailed implementation, and no independent effectiveness benchmark. Cato's claim of protection against previously unseen attacks should be read as the product's intended security capability, not a guarantee that every prompt-injection attempt will be stopped.
The operational question is whether a security layer has enough session and action context to tell the difference between using a document as evidence and treating that document as authority. AgentSense aims directly at that line.
What is Drift Detection?
Cato also previewed a second capability, AgentSense Drift Detection, described in forward-looking terms. The planned feature will evaluate whether an agent's actions continue to serve the original task as a workflow progresses.
The two features attack related but distinct problems. Untrusted Influence Detection asks whether outside content is steering an action. Drift Detection is intended to ask whether the action still belongs to the objective — even if the individual step is technically permitted.
A workflow can gradually accumulate steps that no longer advance its purpose. Detecting that departure requires maintaining a view of the authorized objective across the entire session, rather than judging every tool call in isolation. The release does not specify a separate availability date for Drift Detection.
Where does it fit in Cato's stack?
Cato's agentic AI security offering already covers several layers:
- Discovering agents and Model Context Protocol servers
- Profiling tool calls
- Enforcing policies on agent interactions
The broader Cato platform combines networking, access, and security context. The company also separates security for employee AI use from protection for applications. Its AI Security for Applications offering includes runtime visibility, prompt and response inspection, and configurable protections for deployed AI applications.
That architecture provides the backdrop for AgentSense's focus on the integrity of actions taken during a workflow. In practice, controlling AI use involves several related decisions: which services an organization permits, which data may reach them, and what an agent may do with the information it encounters. AgentSense adds a behavioral claim to that stack — identifying when untrusted content is driving the action itself.
Why prompt injection research proves the point
Cato's own research record shows why this boundary matters.
Its EchoLeak research described an attack chain in which externally supplied content could cause Microsoft 365 Copilot to expose information from its context. The significance for agent security is the transition from reading outside material to acting on instructions embedded in it.
Its CurXecute research examined a different outcome: prompt injection through external content could lead a coding agent to modify an MCP configuration and trigger command execution. Cato's write-up says Cursor fixed that vulnerability in version 1.3. The example illustrates a failure mode; it should not be read as a claim that the patched issue remains open.
Both cases show why content provenance matters once an assistant can reach sensitive data or execution tools. An interaction that begins as information retrieval can become a security event when retrieved text changes the workflow's control flow.
What are the challenge rules?
The AgentSense Challenge invites participants to attempt adversarial workflow redirection against a protected agent. The first three eligible successful participants will each receive $10,000, according to the announcement.
One caveat: the supplied release does not include the final challenge URL or its detailed rules. Eligibility and testing boundaries will depend on the published program terms.
For enterprises expanding agent autonomy, AgentSense addresses a concrete need — preserving the connection between the task a user authorizes and the actions an AI system ultimately executes. The technology's value will depend on how consistently it can enforce that boundary while allowing legitimate work to continue, and the $10,000 challenge is Cato's bet that outside researchers will struggle to prove otherwise.
Source: unite.ai
More from Elena Vasquez
Show full bio
Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.
708 articles
Related articles
- OpenAI adds Lockdown Mode and risk labels to ChatGPT
- Five Agent Vulnerabilities in Five Months Expose Structural Flaw in MCP
- Airlock Digital adds action-level controls for Codex and Claude
- OpenAI Launches Safety Bug Bounty to Pay for AI Abuse Findings
- AI Agents Move Onto Windows PCs, and Permissions Become the New Battleground