Decoding the world of cybersecurity

MPs call for statutory AI safety regulator

A Commons committee has called for enforceable AI safety legislation, a statutory regulator, and mandatory incident reporting, arguing that technical evaluations alone cannot provide adequate public accountability.

MPs call for statutory AI safety regulator
Summary
  • MPs identified ten gaps in the UK's existing AI safety and security arrangements.
  • The committee recommends a statutory regulator with powers distinct from the AI Security Institute.
  • Proposed measures include independent verification, intervention powers, incident reporting, and clearer liability.

Britain should establish a statutory artificial intelligence regulator with enforceable powers to investigate risks and intervene against dangerous systems, according to MPs who say existing safeguards leave significant gaps in accountability.

The House of Commons Business, Innovation, Science and Trade Committee published its report, Guardrails for growth: AI safety and security, on 18 November 2026. The committee identified ten gaps in the UK’s current approach and recommended legislation that would place clearer legal obligations on developers and other organisations responsible for advanced AI systems.

The proposed framework would complement the technical work of the AI Security Institute (AISI) with a separate regulator able to require evidence, enforce safety obligations and intervene when risks cannot be adequately controlled.

The committee gathered evidence from researchers, former industry employees and representatives of major AI developers, including OpenAI, Google DeepMind, Anthropic and Meta.

Testing and enforcement require different powers

AISI conducts technical research and evaluations intended to improve understanding of AI system capabilities and potential hazards. Its expertise can identify risks, but scientific assessment does not itself determine which risks society should accept or establish enforceable obligations on companies.

The committee argues that decisions of that kind belong within a framework accountable to ministers and Parliament. It recommends a statutory regulator with secure, multi-year funding and a clearly defined relationship with AISI.

Proposed powers include requiring independent verification of developers’ safety claims, pausing high-risk training or testing, delaying or prohibiting deployment and suspending services where identified risks cannot be controlled.

The report also calls for mandatory incident reporting, protected whistleblowing arrangements and greater clarity over liability when AI systems cause harm.

These recommendations are not existing legal duties. The government would need to determine the scope of any legislation, introduce the relevant measures and secure parliamentary approval before the proposed framework could operate.

Establishing a regulator would also require decisions about its relationship with existing sector authorities. Financial services, healthcare, communications and other regulated industries already have supervisory arrangements that may apply to activities involving AI.

Responsibility differs across the AI supply chain

The committee’s recommendations address a problem that extends beyond the security of individual models. Developers, technology providers and organisations deploying AI systems can exercise different levels of control over how those systems behave.

A model developer may determine aspects of training, evaluation and release. A cloud provider may operate the infrastructure on which a model runs, while an organisation integrating the model into a business process determines what data it can access and which actions it can perform.

Those responsibilities can overlap without being identical. A vulnerability in model infrastructure may arise from implementation decisions made by the provider, while excessive permissions granted to an automated agent may be determined by the organisation deploying it.

Clearer legal accountability would therefore require a framework that reflects the role of each participant rather than treating every AI-related failure as the responsibility of the original model developer.

Independent assurance is another consideration. Claims about a system’s safety may depend on the evaluation methods used, the conditions tested and the extent to which results remain valid after a model is modified or connected to additional tools.

A regulator able to require evidence could strengthen oversight, but the effectiveness of those powers would depend on technical expertise, access to relevant information and sufficiently precise legal standards.

International development complicates UK oversight

Many widely deployed AI models are developed and operated by companies headquartered outside the United Kingdom. Their use within British organisations creates questions about the reach of domestic regulatory requirements and the responsibilities of overseas providers.

The committee recommends work towards an international convention on AI safety and security, potentially using Britain’s G20 presidency in 2027 to advance common expectations.

International cooperation could support information sharing, incident reporting and comparable assessment standards. However, it would not remove the need for domestic legislation specifying the powers and responsibilities of UK authorities.

The report also links safety regulation with the economic development of AI. Clear requirements may improve confidence among organisations considering adoption, while poorly designed obligations could introduce costs or uncertainty without producing equivalent protection.

The committee wants the government to explain how existing, proposed and new legislation would interact. That would require identifying gaps rather than assuming every AI risk needs an entirely separate statutory regime.

The government’s response will determine whether these recommendations develop into a formal legislative programme. For now, the committee has made the case for separating technical evaluation from statutory enforcement while requiring stronger evidence of how advanced AI systems are managed.

×