UK AI Security Institute

A research body, not a regulator

The UK AI Security Institute is a research organisation inside the British government. It was established in November 2023, within what was then the Department for Science, Innovation and Technology^[1]. Its purpose was defined fairly narrowly and technically: to conduct research on the most powerful foundation models, to test how well safety guardrails actually work, and to inform governments about AI safety^[1]. Foundation models here means the large, general-purpose models that many other systems are built on; guardrails means the filters and restrictions placed around a model to stop it producing harmful output.

The institute is not a regulator. It has no regulatory authority and cannot demand access to foundation models, either before or after they are deployed, nor can it prevent a model from being released^[1]. Its arrangements with developers, which give it access to frontier models for pre-deployment testing, are voluntary^[1]. Its influence therefore rests on technical work and on developers choosing to cooperate, rather than on legal power.

The institute has since moved from the Department for Science, Innovation and Technology to the Cabinet Office^[1].

From safety to security

When it was created, the body was called the AI Safety Institute. In February 2025 it was renamed the AI Security Institute, with a renewed sole focus on the risks AI poses to national security and serious crime^[1]. The Secretary of State announced the renaming at the Munich Security Conference on 14 February 2025^[2].

The change was not only a matter of wording. The renaming removed references to algorithmic bias from the institute's stated agenda^[1], narrowing the set of harms it describes itself as studying towards a security frame: hostile use, crime, and dangers to the state.

Working with the rest of government

The institute does not do this work alone. A new criminal misuse team was launched to research AI-enabled crimes such as fraud, working jointly with the Home Office^[3]. Its scientists work with the Defence Science and Technology Laboratory to assess the dual-use scientific capabilities of advanced AI^[2], meaning the scientific abilities of advanced models that could serve harmful purposes as well as beneficial ones. The institute also works with the Laboratory for AI Security Research and the National Cyber Security Centre^[2].

These partnerships matter because the questions the institute is asked to answer are rarely purely technical. Judging whether a model could help someone commit fraud, or whether a scientific capability is dangerous in the wrong hands, requires expertise that sits in other departments and agencies. The institute's role is to bring model testing together with that expertise.

Where it fits

The institute occupies a specific gap. Governments need someone who can actually test powerful models and interpret the results, rather than relying only on what developers say about their own systems^[1]. Its remit, its access and its audience are all governmental: it tests guardrails and informs governments, but it does not write rules.

That combination — real technical capacity paired with voluntary access and no enforcement powers — makes the institute a particular kind of AI governance arrangement, one built on research and cooperation rather than binding regulation^[1]. Its history also traces a shift in how AI risk has been framed in UK policy: the same organisation, created in 2023 as a safety institute, was reoriented in 2025 towards national security and serious crime, and its stated agenda no longer listed algorithmic bias^[1].