Safety & Security

Cato’s AgentSense cuts AI agent attack success rate to 0.05%

Cato says AgentSense cut Agent Dojo attack success from about 19% to 0.05%, with under 3% added latency, while blocking risky actions driven by untrusted content.

Cato Networks launches AgentSense to block harmful AI agent actions
Cato Networks launches AgentSense to block harmful AI agent actionsPixelkult / pixabay
By Rebecca Stone5 min read

Why it matters

  • AgentSense reduced the Agent Dojo attack success rate from about 19% to 0.05% in Cato’s testing.
  • Cato reported less than 3% added latency with AgentSense.
  • The initial release uses Untrusted Influence Detection to trace untrusted content before an action executes.
  • AgentSense includes three built-in rules and is available with Cato AI Usage licenses at no additional charge.
  • The first three eligible public challenge participants will each receive $10,000.

Cato Networks said AgentSense reduced the attack success rate on the Agent Dojo benchmark from about 19% without additional protections to 0.05%, while adding less than 3% latency.

The company launched the security capability this week inside its cloud-based secure access service edge (SASE) platform. According to Cato, AgentSense blocks an AI agent’s proposed actions when untrusted content appears to influence them.

The launch addresses indirect prompt injection, an attack pattern in which instructions embedded in a webpage, email, document, or tool response can redirect an agent that is already doing a legitimate task. According to Cato, the redirected agent could expose sensitive information or make an unauthorized change even when the user’s original request is routine.

“Giving agents greater autonomy requires security that can intervene before their actions cause harm,” Shlomo Kramer, CEO and co-founder of Cato, said in a press release. “Our vision for AgentSense is to give enterprises the confidence to put agents in charge of critical business processes while retaining control over how they operate.”

Kramer’s framing places AgentSense between access grants and execution. A valid credential can open a legitimate door, but Cato says an agent can still act on harmful instructions it discovers after access has been granted. That is the control problem AgentSense is intended to address.

For enterprises, the policy stakes are whether a permission granted for one task can be insulated from instructions encountered later in the same session.

The initial release targets indirect prompt injection, according to Cato. In that pattern, instructions arrive through a webpage, email, document, or response from a tool the agent uses. The original request may remain benign while the embedded text attempts to redirect the agent.

Kramer said agents “can have exactly the right permissions and still do bad stuff” during the launch briefing. The warning addresses a specific weakness in permission-only controls: access rights do not reveal what an agent will do after it reads untrusted material.

How does Cato evaluate an agent’s action?

Cato developed Untrusted Influence Detection for this problem. According to the company, the technology distinguishes trusted instructions from untrusted content within an AI session. It then traces how that content influences a proposed action.

AgentSense can block the action before it executes, according to Cato. The company says the approach can catch attempts outside known attack patterns. That gives the product a different proposition from a control that recognizes only previously documented attacks.

AgentSense works across AI agents, and administrators can enable it with a toggle in the Cato SASE platform, according to Cato. Because the capability sits within the access platform, the company is presenting it as a broad deployment point rather than a control tied to one named agent.

Cato says the platform ships with three built-in rules:

  • Untrusted input steering an agent toward a high-risk operation.
  • Sensitive information about to be exposed externally.
  • Changes to critical resources driven by untrusted input.

In Cato’s taxonomy, the rules cover three outcomes: a dangerous operation, an external disclosure, and a modification of a consequential resource. Each rule begins with untrusted influence, allowing the platform to intervene before a proposed action executes.

How did the Claude Code test work?

Cato described a developer asking Claude Code to set up a test environment. During the task, the agent gathered information from a webpage that instructed it to modify a sensitive local file.

AgentSense traced the proposed file change to the webpage and blocked it before Claude Code could alter the file. According to Cato, the setup therefore remained incomplete. The company presented the case as a successful interception, not a completed test setup.

The example preserves the distinction between refusing the developer’s request and refusing an instruction found inside external material. According to Cato, the agent retained its legitimate task context while AgentSense stopped the untrusted direction.

What do the benchmark numbers show?

The headline figures come from Cato’s own testing on Agent Dojo. Cato reported an attack success rate of about 19% without additional protections and 0.05% with AgentSense. The company also reported less than 3% added latency.

These are vendor-reported results, not an independent assessment of the product. Cato says its test used the Agent Dojo benchmark. The public challenge gives participants a separate way to examine a workflow protected by AgentSense.

The latency figure frames the launch around operational cost as well as prevention. Cato reports keeping added delay below 3% in its Agent Dojo test, making the performance claim part of the same benchmark result as the attack-rate reduction.

What is available now, and what comes next?

AgentSense is available to Cato customers and carries no additional charge when included with Cato AI Usage licenses, according to the company. The current release therefore centers on prompt-injection defenses and platform activation, while several planned controls remain in development.

Cato also plans a Drift Detection capability. According to the company, it will check whether an agent’s actions still serve the user’s original task. Kramer said during the launch briefing that the system should block actions that drift from the original purpose assigned to the agent.

That planned control addresses task continuity rather than only the trustworthiness of an input. A session could contain no single malicious instruction and still wander from the requested job, according to Cato’s described roadmap.

Cato also plans controls over which resources an agent can reach, including the data it can access. The company has not placed those restrictions in the initial release, so AgentSense’s present configuration remains focused on the three built-in rules and Untrusted Influence Detection.

Cato opened a public challenge to test an AI agent workflow protected by AgentSense. According to the company, the first three eligible participants to complete the challenge will each receive $10,000. Cato says it will use participants’ findings to evaluate and improve the guardrail.

The challenge puts a practical test around the product’s main claim: interventions can stop actions shaped by untrusted content before execution. Cato will use whatever participants find to assess the protection, which makes the results a direct input to its next refinement.

Cato has set the next test: challenge participants will try the protected workflow, and the company says their findings will feed its evaluation and improvement of the guardrail.

Source: networkworld.com

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

802 articles

Related articles

  1. Cato Networks Launches AgentSense to Block AI Agent Attacks
  2. Mitiga Launches Agentless Security Platform for Enterprise AI Agents
  3. Obsidian Security Launches Runtime Policy Engine for AI Agent Actions
  4. Onyx and Glean integrate controls for enterprise AI agents
  5. OpenAI adds Lockdown Mode and risk labels to ChatGPT

« Previous articleNext article »