Cato Networks Bets $30,000 That Its AgentSense Can Stop AI Agent Hijacks
Cato Networks launches AgentSense to block AI-agent hijacks via indirect prompt injection, backed by a $30,000 challenge to researchers who can break it — but no independent proof yet.
By Olivia Hart
4 min read

What's News
- Cato Networks unveiled AgentSense, a runtime security capability that aims to stop harmful AI-agent actions before execution.
- The AgentSense Challenge offers $10,000 each to the first three eligible participants who successfully redirect a protected workflow.
- AgentSense is included at no additional charge within Cato AI Usage licenses.
- Cato's CurXecute research led Cursor to fix a prompt-injection flaw in version 1.3.
- Drift Detection is a planned future capability with no announced availability date.
By Olivia Hart, Business Bearings
Cato Networks will pay $10,000 each to the first three researchers who can hijack a protected AI-agent workflow — a $30,000 public wager that its new AgentSense technology can stop harmful agent actions before they execute. The company unveiled AgentSense, a runtime security capability built by Cato AI Labs, to counter a risk that conventional access controls do not address: an agent holding perfectly valid credentials can still be manipulated into doing the wrong thing.
Cato says AgentSense is integrated into its platform and ships at no additional charge within Cato AI Usage licenses. That pricing claim, like the security claim, awaits independent verification — the release offers no benchmark proving the product works as advertised.
Why valid permissions are not enough
An enterprise agent typically reads emails, searches documents, consults websites, and calls external tools. Every source widens the agent's information supply — and every source is also a channel an attacker can use.
Indirect prompt injection exploits exactly that boundary. Rather than issuing a command as the authorized user, an attacker hides an instruction inside material the agent encounters mid-task: a document, a ticket, a web page, or a tool response. The embedded text can redirect the workflow, expose information, or trigger an unrelated action.
Checking whether an agent may access an application answers a different question than whether a specific action serves the task the user actually authorized. Cato's own example: an agent may legitimately read business records without being authorized to send them to a destination suggested inside a retrieved document. AgentSense's pitch is behavioral evaluation during execution, with a chance to intervene before a sensitive action runs.
The takeaway: access control governs reach, not intent — and AgentSense targets the gap between the two.
How does the detection actually work?
The announced Untrusted Influence Detection capability 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 the system blocks an action when it detects outside content steering the agent toward an unsafe step.
The design choice matters. An attack does not need to resemble a known malicious prompt. Cato describes AgentSense as assessing influence rather than relying on static classifiers or a catalog of previously observed attack patterns — an approach intended to catch new forms of indirect prompt injection, not just recognizable ones.
Caveats remain. The release calls the technology patented but supplies no patent identifier, no implementation detail, and no independent effectiveness testing. The claim of protection against previously unseen attacks is the product's intended capability, not a guarantee that every injection attempt will fail.
The core technical question is whether a security layer can gather enough session and action context to distinguish a document used as evidence from a document treated as authority. AgentSense aims squarely at that distinction.
What about Drift Detection?
Cato also previewed AgentSense Drift Detection in forward-looking terms. The planned feature will evaluate whether an agent's actions continue to serve the original task as a workflow progresses — a different problem from influence detection. Untrusted Influence Detection asks whether outside content is steering an action; Drift Detection asks whether the action still belongs to the objective even if the step is technically permitted.
Workflows can accumulate steps that no longer advance their purpose, and detecting that departure requires tracking the authorized objective across the whole session. The release gives no availability date for Drift Detection, so treat it as roadmap, not product.
Where does AgentSense sit in Cato's stack?
Cato's existing agentic AI security offering already covers discovering agents and Model Context Protocol servers, profiling tool calls, and enforcing policies on agent interactions. Separately, its AI Security for Applications provides runtime visibility, prompt and response inspection, and configurable protections for deployed AI applications.
AgentSense adds a behavioral layer to that stack: identifying when untrusted content is driving the action itself.
The research behind the claim
Cato's EchoLeak research described an attack chain in which externally supplied content caused Microsoft 365 Copilot to expose information from its context. Its CurXecute research showed prompt injection leading 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. Both cases illustrate 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.
Can researchers break it?
The AgentSense Challenge invites the security community to plant instructions in a protected workflow and steer the agent away from the user's intent. Cato says findings and participant feedback will inform further evaluation and improvement. The supplied release omits the final challenge URL and detailed rules, so eligibility and testing boundaries depend on terms not yet published.
For enterprises expanding agent autonomy, AgentSense targets a concrete need: keeping the actions an AI system executes tethered to the task a user actually authorized. Its value will hinge on how consistently it enforces that boundary — the $30,000 challenge is, for now, the only public test on offer.
Source: unite.ai
More from Olivia Hart
Show full bio
Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,224 articles