NCSC Sets Out Interim Guidance for Managing Agentic AI Cyber Risk

The National Cyber Security Centre has published interim advice on managing the cyber risk posed by agentic AI, following a string of incidents involving AI models and autonomous systems carrying out unsanctioned activity.

Agentic AI systems can automate complex workflows and free staff to focus on higher-value work, but the NCSC is warning that organisations deploying them at pace need to plan carefully for what happens when these systems behave unexpectedly. The guidance follows several recent incidents in which AI models and agentic systems carried out unsanctioned or unintended activity.

The advice is aimed at system designers and operators building environments where AI agents operate with significant autonomy, or who are concerned about agents taking unintended actions based on the instructions, tools or systems available to them. The NCSC stresses that recommendations should be applied proportionately, based on how much autonomy a given system actually needs and the level of risk an organisation is willing to accept.

The guidance sets out seven key considerations for organisations to work through:

• identifying what could go wrong through threat modelling and defining clear "red lines"

• prompting agents carefully, with explicit instructions on what they should and should not do

• setting the right level of human oversight, from human-in-the-loop approval to fully autonomous operation

• controlling the agent's environment with a robust, isolated sandbox

• logging, auditing and monitoring agentic AI activity as part of security operations

• making AI activity easy to attribute if it interacts with third-party systems

• maintaining the ability to shut down agentic AI activity immediately if an incident occurs

The NCSC notes that model-level safeguards should not be relied on alone, as they can be bypassed or may not provide adequate protection in higher-risk environments. This interim advice will be superseded by formal guidance the NCSC is developing with partners as the evidence base matures.

Organisations working with or advising on agentic AI deployments can read the full guidance and further reading recommendations on the NCSC website.