UK’s trusted IT infrastructure partner since 2003
Servnet
FinanceToolsConfiguratorGet in Touch
Cyber security

Cato launches AgentSense to block harmful AI-agent actions

London · Servnet News Desk · IT infrastructure analysis5 min read
Share

Cato Networks has built a guardrail called AgentSense into its SASE platform to stop AI agents acting on hidden instructions buried in web pages, emails or documents. The launch highlights why agent autonomy may require enforceable controls alongside existing access controls, such as zero trust network access.

AgentSense impact on attack success rate
20%15%10%5%0%19%Without Controls0.05%With AgentSenseAttack success rate
View the data behind this chart
AgentSense impact on attack success rate
Without ControlsWith AgentSense
Attack success rate19%0.05%

Why AI agents have become a live attack surface

AI agents don't need to be hacked to cause harm — they just need to read the wrong thing. Cato Networks' launch of AgentSense targets exactly this weakness: indirect prompt injection, where instructions hidden in a webpage, email, document or tool response steer an otherwise-authorised agent into exposing data or changing something it shouldn't.

Cato CEO Shlomo Kramer framed the problem bluntly during the launch briefing, noting that agents "can have exactly the right permissions and still do bad stuff." That distinction matters for UK buyers: this isn't a permissions failure that identity management alone fixes, it's a behavioural failure that requires inline inspection of what an agent is about to do, not just what it's allowed to do.

The risk is not theoretical. The Register reported, citing Google Threat Intelligence, that a suspected Russian-speaking operator used an AI agent to turn a published security patch into a working exploit without human intervention, then used it to steal data from victims. Separately, BleepingComputer reported that Microsoft's official Agent Framework package was found compromised on PyPI, with malicious versions designed to steal cloud credentials and other sensitive information from developers. Autonomous tooling is already being weaponised on both sides of the fence.

Inside AgentSense: what the control actually does

AgentSense sits inside Cato's cloud-based SASE platform and is enabled with an administrator toggle. Its core mechanism, Untrusted Influence Detection, separates trusted instructions from untrusted content inside an AI session, traces how that content shapes a proposed action, and can block the action before execution — Cato says it can catch attempts that fall outside known attack patterns.

The initial release ships with three built-in rules: untrusted input pushing an agent toward a high-risk operation, sensitive information about to be exposed externally, and changes to critical resources driven by untrusted content. In one documented example, a developer asked Claude Code to set up a test environment; while gathering context, the agent read a webpage carrying instructions to modify a sensitive local file. AgentSense traced the proposed change back to that webpage and blocked it before Claude Code could modify the file.

In Cato's own testing against the Agent Dojo benchmark, the attack success rate dropped from about 19% without additional protections to a Cato-reported 0.05% with AgentSense enabled, with Cato reporting less than 3% added latency in that testing. Cato says AgentSense is available to its customers and included at no additional charge with Cato AI Usage licenses, and the company has opened a public challenge to test an AgentSense-protected workflow, with the first three eligible participants to complete the challenge each receiving $10,000 — a transparency move worth watching as the control matures. Planned additions include Drift Detection, which checks whether an agent's actions still serve its original task, and resource-scoping controls over what data and systems an agent can reach.

Where this lines up with NCSC guidance

UK buyers evaluating AgentSense-class controls shouldn't treat them as a single-vendor novelty — they map closely onto what the National Cyber Security Centre is already advising organisations to do. NCSC guidance recommends running agents in sandboxed environments with both inbound and outbound traffic denied by default and opened only through allowlists, plus the ability to rapidly restrict network access or interrupt communication between an agent and its model-inference backend.

NCSC also pushes for threat modelling before deployment — covering prompt manipulation, misuse of connected tools, and possible escape from the sandbox through logging infrastructure — combined with least privilege, temporary credentials, secure defaults, and continuous behavioural monitoring. A toggle-based blocking layer like AgentSense addresses part of that picture; it doesn't replace the sandboxing, segmentation and human-oversight requirements NCSC sets out separately.

Lateral movement risk in hybrid networks

According to the NCSC's Impact of AI on cyber threat assessment, AI is accelerating reconnaissance and improving precision in spotting vulnerable devices, with less uplift so far in lateral movement compared with reconnaissance, social engineering and data exfiltration. That's a narrow window, not a reason for complacency — an agent with legitimate access to internal systems may already have access paths that could facilitate lateral movement if its instructions are hijacked, subject to segmentation and privilege.

For operators of critical infrastructure, this reinforces older segmentation principles rather than replacing them. Commentary on UK critical-infrastructure design has highlighted that conventional firewalls and VPNs can leave unnecessary return paths into a site, and that hardware-enforced data diodes can remove the ability to send commands back in where facilities only need to export monitoring data. Segmentation that limits an attacker's "time to objective" after a breach, and realistic decoys to divert automated tools, remain relevant countermeasures as agentic workflows spread across hybrid environments. NCSC has separately assessed that AI-powered malware is highly likely to be used in a ransomware attack against UK critical national infrastructure before 2028.

Supply chain is part of the agent attack surface

The compromised Microsoft Agent Framework package on PyPI is a reminder that the agent runtime itself can't be assumed trustworthy. Buyers rolling out agentic AI need to validate third-party frameworks, tools, models and integrations with the same rigour applied to any other software dependency — not treat the orchestration layer as a safe default simply because it comes from a recognised vendor.

Combined with behavioural guardrails like AgentSense, this points to a layered requirement: inspect what an agent is about to do, and separately verify what it's built on. Teams already running managed detection and response should extend that visibility to agent sessions and the packages that support them, rather than leaving agent telemetry as a blind spot.

Agent control maturity: baseline vs guardrails
No AI GovernanceNCSC BaselineAgentSense-ClassSASEPrompt injection defenceNonePartial (sandbox)Blocks before executionOutbound/egress controlOpen by defaultDeny-by-defaultallowlistAction traced & blockedLateral movement riskHighReduced viasegmentationConstrainedby tool scopeSupply-chain validationNot routineRecommendedRequiredalongside controlIncident kill switchManual/slowRapid isolation advisedToggle-based control
View the data behind this chart
Agent control maturity: baseline vs guardrails
No AI GovernanceNCSC BaselineAgentSense-Class SASE
Prompt injection defenceNonePartial (sandbox)Blocks before execution
Outbound/egress controlOpen by defaultDeny-by-default allowlistAction traced & blocked
Lateral movement riskHighReduced via segmentationConstrained by tool scope
Supply-chain validationNot routineRecommendedRequired alongside control
Incident kill switchManual/slowRapid isolation advisedToggle-based control

A practical control checklist for infrastructure buyers

No single product — AgentSense included — covers the full control set NCSC and recent incidents suggest is necessary. Treat AgentSense as representative of a class of controls rather than a finished specification, and build procurement and governance around the following.

  • •Identity and privilege restriction with least-privilege, time-limited credentials for every agent
  • •Sandboxing with deny-by-default egress, opened only through explicit allowlists
  • •Inline action inspection to catch prompt-injection-driven behaviour before execution
  • •Network segmentation and rapid kill switches to contain lateral movement if an agent is subverted
  • •Central telemetry and behavioural monitoring feeding existing security operations
  • •Supply-chain validation of agent frameworks, tools and model integrations
  • •Human approval gates for consequential or irreversible actions
Share
Key takeaways
  • ✓Cato's AgentSense blocks AI agents from acting on hidden instructions found in web content, emails or tool outputs, cutting its Cato-reported Agent Dojo test attack success rate from about 19% without additional protections to 0.05% with AgentSense enabled.
  • ✓NCSC guidance already recommends sandboxing, deny-by-default egress and rapid network isolation for agent deployments — AgentSense-class controls address part of this, not all of it.
  • ✓A Google-reported incident shows AI agents can now turn patches into working exploits without human help, raising the stakes for constraining agent tool use and access.
  • ✓A compromised Microsoft Agent Framework package on PyPI shows agent supply chains need the same validation as any other software dependency.
Frequently asked

FAQs — Cato launches AgentSense to block harmful AI-agent actions

What does Cato's AgentSense actually block?

AgentSense targets indirect prompt injection — instructions hidden in webpages, emails, documents or tool responses that steer an AI agent into unauthorised actions. It traces untrusted content's influence on a proposed action and can block it before execution.

Is AgentSense a replacement for NCSC's agentic AI guidance?

No. NCSC guidance covers sandboxing, deny-by-default network controls, threat modelling and incident response for agent deployments. AgentSense-class behavioural blocking is one layer within that wider framework, not a substitute for it.

How much does AgentSense cost?

Cato says AgentSense is available to its customers and included at no additional charge with Cato AI Usage licenses.

Why does lateral movement matter if AI isn't yet very good at it?

NCSC assesses current AI uplift in lateral movement is lower than in reconnaissance or social engineering, but an agent with legitimate internal access may already have access paths that could facilitate lateral movement if hijacked — making segmentation and access controls essential regardless of AI capability.

Related

Turning this into a buying decision?

One conversation with an engineer who's specced this before. No sales script.

Talk to Servnet →

Talk to a UK specialist

Get expert advice or a no-obligation quote — servers, storage, networking, maintenance, finance and cloud. We reply the same working day.

or call 0800 987 4111