The UK opens the first state AI Safety Institute
Backed by a £300 million compute allocation and chaired by Ian Hogarth, the institute converted a temporary taskforce into a permanent evaluator with formal US and Singapore partnerships.
- Government & policy
- Safety & alignment
- Notable
On the closing day of the Bletchley Park summit, Rishi Sunak formally launched the UK’s AI Safety Institute, converting the Frontier AI Taskforce — a temporary body assembled earlier that year — into a permanent government organisation with a standing mandate to test frontier AI models before and after release. Sunak framed the case for a state evaluator directly: “Until now, the only people testing the safety of new AI models have been the very companies developing them.”
The institute sat within the Department for Science, Innovation and Technology, chaired by Ian Hogarth, who had chaired the Taskforce, with a chief executive still to be recruited. Its launch was backed by a separately announced £300 million expansion of the UK’s AI Research Resource, including priority access to two new supercomputers, Isambard-AI in Bristol and Dawn in Cambridge, intended to give the institute the compute needed to actually run large-scale evaluations rather than rely on lab-supplied infrastructure. The government said it had formalised partnerships with the US’s newly announced counterpart institute and with Singapore’s Infocomm Media Development Authority to share testing capability and methodology.
Unlike the Bletchley Declaration signed the same week, which was a diplomatic statement with no enforcement mechanism, the AI Safety Institute was a standing technical body with its own budget, staff and evaluation programme — the summit’s most concrete institutional legacy. The announcement noted several major labs had committed to giving it pre-deployment access to their models, though the initial press materials did not specify precise obligations or which labs beyond the general expressions of support from OpenAI, Google DeepMind, Anthropic, Meta and Microsoft.
The institute was explicitly framed as an evaluator, not a regulator — its findings would inform UK and international policymaking rather than carry independent legal force. It was renamed the AI Security Institute in 2025 as its remit shifted further toward national-security-relevant risks, and the US launched its own institute the same week, beginning a pattern of parallel, loosely coordinated national AI-safety evaluation bodies rather than a single international one.