The AI Security Institute’s website still describes it as “a research organisation within the UK government’s Department for Science, Innovation and Technology.” That department was closed on 21 July 2026. The page was live and unchanged this week.
It is a small thing, and it is also an accurate description of what happened. Britain moved the body that tests frontier AI models from a sponsoring department into the centre of government, gave artificial intelligence a seat at the Cabinet table, and changed nothing whatsoever about what any AI system in the country is legally required to do.
Where everything went
Prime Minister Andy Burnham abolished the Department for Science, Innovation and Technology on 21 July and split its work three ways. Science and innovation went to a new Department for Business, Innovation, Science and Trade. The digital brief, including the Government Digital Service, telecoms and digital identity, went to an enlarged Department for Digital, Culture, Media and Sport. AI strategy, public-sector AI adoption and the AI Security Institute went to the Cabinet Office, inside the Office for the Prime Minister and the Cabinet.
Kanishka Narayan, previously the AI minister at DSIT, became Minister of State for Artificial Intelligence across the Cabinet Office and the new business department, attending Cabinet.
Three days later the government appointed Lord Vallance to chair a new Prime Minister’s AI Taskforce, sitting in the same office alongside No 10 and core Cabinet Office functions. The taskforce reports to Cabinet Secretary Antonia Romeo and to Narayan, with Vallance reporting directly to the Prime Minister. The announcement confirmed that “responsibility for the existing AI Security Institute will move to the Office for the Prime Minister and the Cabinet.”
“AI is rapidly changing our world and we need to make sure it works for everyone,” Burnham said. Vallance said the potential “to transform public services, drive growth and improve people’s lives is immense.”
The audience changed, not the work
The institute’s technical programme is unaffected by any of this. It runs pre-deployment evaluations of frontier models under agreements with the labs, publishes methodology, and funds outside research.
Its Frontier AI Trends Report gives a sense of what that produces. Performance in some measured domains is doubling roughly every eight months. On chemistry and biology benchmarks, models now exceed PhD-level expert baselines by as much as 60 per cent. On cyber tasks, a model first completed work rated at expert level, meaning ten or more years of experience, in 2025, having been at apprentice level in 2023. Success on hour-long software engineering tasks rose from under 5 per cent in late 2023 to above 40 per cent. Self-replication success rates went from 5 per cent to 60 per cent over the same period. Universal jailbreaks were found for every system tested, though some providers had raised the expert effort needed to breach safeguards fortyfold in six months.
That is capability measurement. It is not compliance measurement, because there is nothing to comply with. The UK has no AI Act. As of the middle of this year no AI Bill sat before Parliament. Testing happens because companies agree to it.
So the question of where the institute sits is not administrative. Evaluation output is an input to a decision, and the location determines whose decision. Inside DSIT it informed a department writing technology policy and sponsoring regulators. Inside the Office for the Prime Minister and the Cabinet it informs the centre, next to national security assessment and public-service delivery. The same cyber capability finding lands on a different desk and gets used for a different purpose.
Where Britain does legislate, it legislates by framework
The counterexample is in the Lords. The Cyber Security and Resilience (Network and Information Systems) Bill, which updates the 2018 NIS Regulations, reached committee stage on 1 September after clearing the Commons and completing second reading on 14 July. Royal Assent is expected late this year.
The duties are not. The Bill brings managed service providers, data centres and designated critical suppliers into scope for the first time, sets a 24-hour initial incident notification with a fuller report at 72 hours, and carries penalties of up to £17 million or 4 per cent of global turnover. But the thresholds, the definitions and most of the substance sit in secondary legislation that follows a government implementation consultation, with practical effect not expected until around 2028.
The Lords Delegated Powers and Regulatory Reform Committee took a view on that approach in its seventh report of the session, published on 21 July. It drew attention to powers allowing ministers to amend the Bill itself by statutory instrument, including a power over the procedure for issuing codes of practice, recommended that one such power be removed altogether and that another be narrowed so it could not be used to amend Acts passed in future sessions. On an early clause it concluded that ministers were seeking a power before either the problem or the required policy response had been identified, and recommended its removal.
Two instruments, one pattern. Where Britain regulates technology, it passes a frame and fills it later by regulation. Where it does not regulate, it builds institutions and moves them.
What it means for firms
The practical consequence is a split between capability and authority. The bodies that can fine you over an AI system are the sector regulators: the Financial Conduct Authority, which has been explicit that it will supervise AI through existing rules rather than write new ones, plus Ofcom and the Information Commissioner. The body that can characterise what a model is actually able to do now sits with the Prime Minister’s office. No statute connects them.
The Institute for Government, writing on 24 July, warned that the split risks losing focus, expertise and coordination, noting that three departments now have to work out their respective roles and that the change largely reverts to the arrangements that existed before 2023. Machinery-of-government changes are not free.
Three things to watch. Whether an AI Bill appears in the next King’s Speech, which would finally give the evaluations a statutory hook. Whether the institute keeps publishing openly now that it reports through the centre rather than a department. And whether the cyber bill’s secondary legislation slips further, having changed sponsoring department halfway through its passage.
Until then, the most consequential AI oversight capability in British government is a research team with no rulebook, a new address and an out-of-date job description.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.