UK AI Safety Institute Publishes Evaluations of Five Frontier Models
The UK AI Safety Institute published its first evaluations of five frontier AI models, testing for catastrophic risk capabilities — but the regime remains voluntary, and the Institute is seeking a statutory right of access.
The UK AI Safety Institute released its first comparative evaluation of five frontier AI models in early 2026, testing for capabilities in biological, chemical, and radiological risk, as well as offensive cyber, persuasion, and autonomous task completion. The results were shared with the model developers before publication, and in two cases the Institute asked for changes to deployment configurations before public release.
The evaluations were conducted under voluntary agreements with the developers, not statutory authority. The Institute has indicated it is seeking a legislative basis for mandatory pre-deployment access to frontier models.
Why it matters: Pre-deployment evaluation by an independent government body is a meaningful intervention if the results influence how models are released. The voluntary nature of the current regime is the binding constraint: without a statutory right of access, developers can decline or limit cooperation. The UK's move toward mandatory access will be the moment this programme acquires real regulatory weight.
