The UK National Cyber Security Centre (NCSC) has issued interim guidance urging organizations to implement stronger controls for autonomous AI agents. The advice, published on August 20, recommends sandboxing, human oversight, and tightly controlled access to mitigate risks from unintended or malicious agent activity. This follows previous NCSC guidance on securing agentic AI use and comes amidst ongoing development of governance frameworks for increasingly autonomous systems. The NCSC noted that formal guidance is still under development and will eventually supersede the current blog post.
Organizations are advised to first assess the necessary level of autonomy for a system and identify potential failure points before deployment. The NCSC recommends threat-modeling the agent's prompts, tools, networks, and accessible services to determine required controls. It explicitly warns against relying solely on safeguards built into underlying models or agent frameworks, as these can be bypassed or prove insufficient in high-risk scenarios.
For higher-risk deployments, the NCSC suggests running agents in robust sandboxes and restricting their access to only the resources essential for a given task. Network controls should deny connectivity by default, utilizing allowlists or service-aware proxies for necessary connections. The agency also recommends separating agent execution, supporting infrastructure, and inference services where feasible. A key concern highlighted is the potential for agents to discover configuration weaknesses or vulnerabilities in their technical controls, which could lead to sandbox escapes.
Each agent should be assigned a distinct identity with credentials limited to those required for its task, ideally using short-lived credentials. Organizations should consider API keys, OAuth grants, SSH keys, and authenticated sessions as part of an agent's potential "blast radius."
Human oversight is crucial for higher-risk activities, with clear responsibility for agent operations, real-time monitoring, and the ability to intervene in unexpected behavior. Agent activity should be logged and monitored as part of security operations and incident response. The NCSC also emphasizes the importance of being able to immediately halt autonomous activity, including restricting network access and communications with model infrastructure when necessary.
The agency stressed that this advice should evolve with the technology, and organizations should regularly reassess whether the autonomy granted to agents remains proportionate to their risk tolerance. The guidance comes as the cybersecurity community continues to grapple with the implications of advanced AI systems, with other initiatives like the OWASP Agentic AI Security Maturity Framework also emerging to address these challenges.
