SIAgentNews.com
← All news
DevelopingAgentsSecurity

Cato Networks launches AgentSense to block AI agent manipulation

The new security tool aims to stop 'indirect prompt injections' where untrusted content tricks an AI agent into performing unauthorized actions.

· First reported by Unite.AI

$10,000prize for successful adversarial testing

The short version

  • AgentSense identifies when untrusted material steers an AI agent toward unsafe steps during a workflow.
  • The tool distinguishes between using a document as evidence versus treating it as an authority.
  • Cato Networks is offering a $10,000 challenge to researchers to test the tool's defenses.
  • AgentSense is being included at no extra cost within Cato AI Usage licenses.

Why it matters

As companies give AI agents more autonomy to read emails and use tools, they face risks from hidden instructions in external data. AgentSense attempts to ensure that an agent's actions remain aligned with the user's original intent rather than the content it encounters.

The risk of hidden instructions

Enterprises are increasingly deploying AI agents that can read emails, search internal documents, and interact with various websites. While these agents need broad access to be useful, each new source of information creates a potential entry point for manipulation. An attacker can hide malicious instructions within a document, a support ticket, or a web page that the agent is tasked with reviewing.

This specific type of attack is known as an indirect prompt injection. Unlike a direct prompt where a user tells the AI to do something bad, an indirect injection occurs when the AI encounters outside text that attempts to hijack the workflow. The agent might be tricked into exposing private information or triggering an unrelated, harmful action based on the content it just processed.

Cato Networks developed AgentSense to address this specific vulnerability. The company notes that having valid credentials does not guarantee safe behavior because an agent can still be manipulated into doing the wrong thing. AgentSense is designed to monitor these interactions in real time to stop harmful actions before they are actually executed.

How AgentSense detects influence

The core technology behind AgentSense is called Untrusted Influence Detection. This capability works by separating trusted instructions—those provided directly by the authorized user—from untrusted content found during the agent's session. It then traces how that untrusted material affects the agent's subsequent behavior and final outputs.

Cato explains that the system does not rely on static classifiers or a pre-defined catalog of known attack patterns. Instead, it assesses the influence of the content on the agent's logic. This approach is intended to help the system recognize new forms of prompt injections that have not been seen before by security teams.

The primary goal of this detection is to distinguish between using a document as evidence and treating that document as an authority. For example, an agent might be permitted to read a business record to find a fact, but it should not be allowed to follow a command inside that record to send the data to an external destination. AgentSense aims to intervene at that specific moment of decision.

Distinguishing drift from influence

While AgentSense focuses on external influence, Cato Networks also described a separate planned feature called AgentSense Drift Detection. This upcoming capability will check if an agent's moves stay aligned with the initial goal as the process moves forward. It serves as a complement to influence detection rather than a replacement for it.

Drift Detection is intended to ask a different question: does the current move still match the primary aim, even if that specific step is allowed by the system? A complex workflow can gradually accumulate many small steps that eventually lead the agent away from its purpose. This feature would require the system to maintain a continuous view of the authorized objective across the entire session.

By monitoring for drift, the system can identify when a workflow has wandered off track. While influence detection looks for outside content steering the agent, drift detection looks for a general departure from the user's goal. Cato has not yet provided a specific availability date for the Drift Detection feature.

Testing and platform integration

Cato Networks is integrating AgentSense into its existing platform at no additional charge for customers with Cato AI Usage licenses. The company already offers a suite of agentic AI security controls, including tools for discovering agents, profiling tool calls, and enforcing policies on how agents interact with various servers.

To test the effectiveness of the new technology, Cato is launching an adversarial testing challenge. They are offering $10,000 each to the first three eligible participants who can successfully redirect a protected workflow into malicious actions. This challenge is intended to provide feedback and help the company improve the tool's defenses against sophisticated attacks.

The company's research has previously highlighted the dangers of these vulnerabilities. For instance, their EchoLeak research showed how external content could cause a copilot to expose context, while their CurXecute research showed how a coding agent could be tricked into modifying configurations. AgentSense is the company's proactive response to these types of security failures in autonomous workflows.

What to watch

  • Whether the $10,000 challenge reveals common bypasses for the influence detection logic.
  • The official release date and specific rules for the Drift Detection feature.

Try it

Check if your organization's AI agents have permissions to 'write' or 'send' data to external destinations that they only need to 'read' from.

Sources and links

First reported by Unite.AI. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news