SIAgentNews.com
← All news
DevelopingAgentsSecurity

Cato Networks launches AgentSense to stop AI agents from harmful actions

The new security tool monitors AI agents in real time to prevent them from following malicious instructions hidden in external data.

· First reported by PR Newswire

25 percentof enterprise breaches driven by ungoverned AI agent abuse by 2028

The short version

  • AgentSense identifies when an AI agent's actions drift away from its original intended task or purpose.
  • The system blocks actions in real time before they execute, rather than just flagging them after the fact.
  • It specifically targets 'indirect prompt injections' where malicious instructions are hidden in emails, docs, or webpages.
  • Cato Networks is offering a $10,000 challenge for security researchers to try and bypass the new defense.

Why it matters

As companies give AI agents more autonomy to perform business tasks, they face a new risk where agents can be 'hijacked' by external data. AgentSense provides a way for businesses to let agents work independently without losing control over the final outcome.

The rise of autonomous AI risk

Enterprises are increasingly moving toward 'agentic AI,' which refers to AI systems capable of performing multi-step tasks autonomously. While these agents can improve efficiency, they also create a significant security gap because they often lack the ability to distinguish between trusted instructions and untrusted data. As reported by PR Newswire, agents might pick up malicious commands from a webpage, an email, or a document they are processing during a normal workflow.

This creates what security experts call a 'lethal trifecta' where an agent with valid permissions can be redirected to perform unauthorized actions. If an agent has the power to access sensitive data or execute system commands, a single hijacked instruction can lead to a major data breach. Current security models like Zero Trust often fail here because they focus on who has access rather than what an agent is actually doing with that access.

The scale of this problem is growing rapidly as organizations rush to adopt these tools. Gartner reports that 71% of organizations expect to deploy agentic AI within the next 12 months. However, only 21% currently have the necessary governance capabilities in place to manage these risks safely. Without proper oversight, these autonomous systems could become a primary source of enterprise breaches by 2028.

How AgentSense monitors behavior

Cato Networks developed AgentSense to act as a real-time guardrail for these autonomous workflows. Instead of just looking for known attack patterns, the system uses runtime verification to assess an agent's actions as they happen. It specifically looks for 'drift,' which occurs when an agent's steps begin to move away from the original goal the user assigned to it.

One core component of this technology is Untrusted Influence Detection. This feature traces how specific pieces of data influence an agent's behavior during a live session. If the system detects that untrusted content—such as a hidden command in a ticket—is steering the agent toward an unsafe action, it blocks that action before the agent can execute it. This allows the system to stop 'zero-day' vulnerabilities that have never been seen before.

The tool works by separating trusted instructions from the untrusted content the agent encounters. By maintaining this distinction, AgentSense ensures that the agent remains focused on its primary objective. It provides a layer of protection that evaluates whether a series of individually permitted actions are being combined to produce an unauthorized or harmful outcome for the business.

Defending against prompt injections

A major focus of AgentSense is preventing indirect prompt injections. This is a type of attack where a hacker hides malicious instructions in a place where an AI agent is likely to look, such as a public webpage or a shared document. When the agent reads that content, it may follow the hidden instructions instead of the user's original request. This can lead to the agent leaking data or manipulating business processes without the user's knowledge.

Cato Networks highlights several notable examples of these vulnerabilities, including DuneSlide, EchoLeak, and CurXecute. These cases demonstrate why it is critical for enterprises to have a system that can verify the intent of an AI agent. AgentSense is designed to preserve trust by ensuring that even if an agent encounters content controlled by an attacker, it cannot be forced to carry out malicious actions.

By providing end-to-end visibility into both human and AI interactions, Cato aims to give companies the confidence to use agents for critical processes. The goal is to allow agents to have greater autonomy while maintaining the same level of control that companies currently have over financial transactions. This balance is necessary for organizations that want to scale AI without exposing themselves to unpredictable behavior.

Testing and deployment plans

To prove the effectiveness of the new technology, Cato Networks has launched the AgentSense Challenge. This initiative invites the security community to attempt to hack an AI agent workflow protected by the new system. Participants are tasked with planting instructions to steer the agent away from its intended goal and into performing a malicious action. The first three people who successfully complete the challenge will receive $10,000 USD.

This challenge serves a dual purpose for the company. Beyond rewarding researchers, it provides a way to perform adversarial testing to find and fix weaknesses in the software. Cato plans to use the findings and feedback from these participants to strengthen the AgentSense capabilities. This iterative process helps ensure the tool can handle the evolving tactics used by attackers to manipulate AI models.

For existing customers, AgentSense is already integrated into the Cato platform. It will be included at no extra cost for those who have Cato AI Usage licenses. This move reflects the company's strategy to provide a converged network and security cloud specifically built for the AI era. By making these tools accessible, they hope to help thousands of organizations secure their AI adoption as the technology becomes more common.

What to watch

  • Whether 25% of enterprise breaches will actually be driven by AI agents by 2028.
  • How other security companies respond to Cato's runtime verification approach.

Try it

Check out the Cato Networks website to see the specific details of the AgentSense Challenge and the $10,000 prize.

Sources and links

First reported by PR Newswire. 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