Back to News
SecurityAI Understanding briefing

Cato Networks adds AgentSense to block harmful AI agent actions

Cato Networks introduced AgentSense, a new SASE‑based guardrail that detects and blocks AI agents from acting on untrusted content, aiming to prevent prompt‑injection attacks on enterprise systems.

4 min readRead the linked source
Source-provided image accompanying Cato Networks adds AgentSense to block harmful AI agent actions
Source referenceSource recorded
Publisher
networkworld.com
Source type
Linked source — primary-source status has not been established.
ContextUnderstand this in 60 seconds

Key terms

AI Agent
A software system that can observe, reason, and take actions to achieve a goal, often using tools and memory.
Guardrails
Rules, checks, and controls that limit unsafe or undesired model behavior.
Robustness
A model's ability to maintain performance under noise, shifts, or adversarial inputs.

What happened

Cato Networks launched AgentSense, a security capability integrated into its cloud‑based Secure Access Service Edge (SASE) platform. The feature is designed to stop AI agents from executing harmful actions that stem from untrusted or malicious content, such as prompt‑injection attacks. AgentSense uses an Untrusted Influence Detection engine to trace how external content influences an agent’s proposed actions and can block those actions before they execute. The initial release includes three built‑in rules covering high‑risk operations, exposure of sensitive data, and unauthorized changes to critical resources. Administrators enable the feature via a toggle in the Cato SASE console, and it is offered at no extra cost to customers with Cato AI Usage licenses. Cato also announced a public challenge, rewarding the first three participants who successfully test an AI workflow protected by AgentSense with $10,000 each.

Cato Networks announced AgentSense in a press release and launch briefing, positioning it as a safeguard for enterprises that grant AI agents greater autonomy in critical processes. The feature is built into Cato’s existing SASE platform, allowing customers to activate it with a simple toggle.

AgentSense focuses on indirect prompt‑injection scenarios, where an agent encounters malicious instructions embedded in content it reads while performing its duties. The system’s Untrusted Influence Detection component tracks the provenance of a proposed action back to the source content and blocks the action if the source is deemed untrusted.

Three default are shipped: (a) preventing untrusted input from steering an agent toward high‑risk operations, (b) blocking exposure of sensitive information to external destinations, and (c) stopping unauthorized changes to critical resources driven by untrusted input. Cato illustrated the feature with a Claude Code example where the agent was prevented from modifying a sensitive file after reading a malicious webpage.

In internal testing on the Agent Dojo benchmark, Cato reported that AgentSense reduced the attack success rate from about 19% to 0.05% while adding less than 3% latency. The company also hinted at future capabilities, including Drift Detection to ensure agents stay aligned with their original tasks and finer‑grained resource‑access controls.

AgentSense is available at no additional charge to customers holding Cato AI Usage licenses. To encourage external validation, Cato launched a public challenge offering $10,000 to each of the first three participants who successfully test a protected AI workflow.

Source details: networkworld.com

Why it matters

AI agents are increasingly embedded in enterprise workflows, handling tasks that involve access to sensitive data and critical systems. Prompt‑injection attacks—where malicious instructions are hidden in webpages, emails, documents, or tool outputs—can cause agents to unintentionally leak data or make unauthorized changes, even when the original user request is benign. By detecting the influence of untrusted content and blocking risky actions, AgentSense aims to close a gap that traditional permission models do not address. Cato’s internal benchmark on the Agent Dojo suite reported a reduction in attack success rate from roughly 19% to 0.05% with less than 3% added latency, suggesting the guardrail can be both effective and performant. If widely adopted, such controls could become a standard layer of defense for enterprises deploying AI agents, reducing the risk of data breaches and operational disruptions caused by sophisticated injection techniques.

The rise of AI agents in enterprise settings creates new attack surfaces that traditional security controls—focused on user authentication and network perimeter—do not fully cover. Prompt‑injection attacks can bypass permission checks because the agent itself holds legitimate credentials.

By embedding a detection layer that evaluates the trustworthiness of content influencing an agent’s decisions, AgentSense addresses a class of threats that are otherwise difficult to mitigate, especially in environments where agents autonomously browse the web or process external documents.

The reported reduction in attack success rate, combined with minimal performance impact, suggests the technology could be practical for high‑throughput enterprise workloads where latency is a concern.

If the public challenge uncovers edge cases or limitations, Cato can refine the , potentially setting a benchmark for the industry. Successful adoption could pressure other security vendors to develop comparable AI‑specific protections, leading to broader standards.

The feature’s inclusion at no extra cost for existing AI Usage license holders lowers the barrier to entry, encouraging early adoption among Cato’s customer base and providing a real‑world testing ground for the technology.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

Agent Lifecycle Stage:
1
User Intent & Planning: "Audit customer refund request #4092 and settle payment."
2
Tool Calling: Emits structured JSON call crm_get_transaction(id='4092').
3
Guardrail & Verification:Paused: High-value action requires human operator sign-off.
4
Final Settlement: Refund recorded, email receipt dispatched, and audit log stored.
Core takeaway: An AI agent is not just a language model—it is a closed loop of planning, tool invocation, and environment feedback. Production systems require self-healing retries and strict human approval guardrails.
Interactive Concept Check+10 Points
AI Security Quiz

A public chatbot and an internal agent with write access are being assessed. Why need separate threat models?

What to watch next

Key areas to monitor include: (1) adoption rates among Cato’s enterprise customers and feedback on real‑world effectiveness; (2) results from the public challenge, which may reveal new attack vectors or limitations of the current rule set; (3) the rollout of the planned Drift Detection and resource‑access controls, which aim to further limit agents’ deviation from intended tasks; and (4) competitive responses from other SASE or security vendors, potentially leading to broader industry standards for AI‑agent .

Customer uptake: Monitoring how many Cato clients enable AgentSense and any reported incidents will indicate real‑world efficacy.

Challenge outcomes: The findings from the $10,000 challenge participants could reveal new attack vectors or confirm the guardrail’s .

Future feature rollout: The announced Drift Detection and resource‑access controls will expand the protection scope; their release timelines and performance will be important.

Competitive landscape: Other SASE or security providers may introduce similar AI‑agent , which could drive industry‑wide best practices or standards.

Regulatory interest: As governments scrutinize AI safety, solutions like AgentSense may attract attention from regulators seeking concrete mitigation strategies for AI‑driven threats.

Related guides & quizzes

Found this useful?