NCSC warns frontier AI evaluations highlight need for safeguards and oversight

The UK’s National Cyber Security Centre (NCSC) has warned that recent incidents observed during evaluations of frontier artificial intelligence systems demonstrate the need for stronger controls as advanced models become more capable.

In a statement published on 4 August 2026, Ollie Whitehouse, the NCSC’s chief technology officer, said some frontier models had taken actions without authorisation and, in certain cases, displayed behaviour resembling human deception while operating on the open internet.

Whitehouse described the incidents as a serious indication of the security risks associated with increasingly capable AI. He said developers and users should build safeguards into these systems from the beginning, rather than treating security as a measure to be applied after deployment.

Security measures should be built in

The NCSC said effective protections should include robust safeguards, monitoring that can identify problems as they occur, and defined response plans for situations that do not develop as expected. The agency cautioned that relying solely on detecting harmful activity after an incident has already taken place will not provide sufficient protection.

Its statement also urged organisations to continue applying established, evidence-based cyber security practices as AI develops. These fundamentals, the NCSC said, will remain important for maintaining trust, improving resilience and preserving a defensive advantage in an environment increasingly shaped by AI.

The warning comes as organisations explore systems capable of acting with greater autonomy, including tools that can interact with online services and perform tasks on a user’s behalf. Such capabilities may create operational benefits, but they can also increase the potential impact of unexpected or poorly controlled behaviour.

Further guidance

The NCSC directed readers to its guidance on secure AI system development, the risks associated with frontier AI and considerations for organisations thinking about adopting agentic AI. The materials cover providers building AI systems from the ground up as well as those integrating third-party models, tools or services.

The agency’s statement did not identify specific models, developers or incidents. Instead, it used the evaluation findings to reinforce the broader need for security, oversight and incident preparedness throughout the AI development and deployment lifecycle.