AISI has no statutory powers while the Frontier AI Bill slips indefinitely

AISI's model access rests on voluntary agreements with (mostly US) labs; it cannot compel pre-deployment testing, information disclosure or corrective action. The promised legislation ('binding regulation on the handful of companies developing the most powerful AI models') has been delayed since 2024: first to fold in copyright, then displaced by the October 2025 regulation blueprint; as of mid-2026 no AI bill is before Parliament, and the Technology Secretary has signalled minimal new regulation. CLTR and GovAI have detailed statutory-footing proposals; nothing has moved.

Voluntary access is discretionary and revocable: exactly the failure mode the Europe-2031 scenario stresses. AISI's own trends report shows agent task complexity doubling every ~8 months; the window in which the UK can set enforceable terms with labs it depends on is closing.

A Frontier AI Act putting AISI on statutory footing: compelled pre-deployment access above capability/compute thresholds, incident and safety-case disclosure duties, safe harbours for cooperating developers, plus explicit reassignment of the societal-impacts remit dropped in the February 2025 rename.

// State-led: Instrument: a Frontier AI Act putting AISI on statutory footing.

Frontier-model testing rests on revocable voluntary access that decays as commercial stakes rise; detailed statutory proposals exist yet no bill is before Parliament and the window is closing.