NCSC outlines safeguards for managing agentic AI cyber risk

The UK’s National Cyber Security Centre (NCSC) is urging organisations to apply layered safeguards when deploying agentic artificial intelligence systems that can make decisions and perform actions with limited human intervention.

In interim guidance, the agency says autonomous AI can improve productivity by automating complex workflows, but warns that failures may lead systems to act outside their intended scope, access inappropriate information or carry out unauthorised tasks. The advice is aimed at designers and operators building environments in which AI agents can interact with tools, data and business systems.

Match autonomy to risk

The NCSC recommends deciding first how much autonomy is genuinely required. An agent that makes recommendations or performs tightly restricted, low-impact tasks presents a different risk profile from one that can modify production systems or make high-consequence decisions without approval. As autonomy and potential impact increase, organisations should apply stronger controls and define the outcomes they are not prepared to accept.

Organisations should also assess the protections supplied by the underlying model, inference service and agent framework. Built-in safeguards may help prevent unwanted behaviour, but they can be bypassed or prove inadequate in higher-risk settings. Model-level protections should therefore be treated as a baseline, not a complete security strategy.

Layered operational controls

The guidance highlights threat modelling before deployment, including documentation of the agent’s authorised activities and explicit boundaries. Carefully designed prompts, appropriate human approval points and additional technical safeguards—such as classifiers or deterministic checks—can help prevent unsafe actions.

A robust sandbox is recommended to limit the agent’s access to networks, credentials, files and other systems. Organisations should log and audit activity, monitor agents through existing security operations processes, and make actions attributable to a specific agent, user or service account. These measures can support investigation and response when behaviour deviates from expectations.

The NCSC also advises maintaining an emergency mechanism to suspend or disable an agent. Shutdown procedures should be tested and designed so they remain effective if the system is malfunctioning or acting unexpectedly.

The agency describes the recommendations as interim advice while it develops more formal guidance with industry and research partners. It expects effective practice to change as evidence about autonomous AI systems grows.