The UK AI Security Institute sets out eight priorities for technical best practice at frontier labs, including reporting incidents caused by AI agents

The UK AI Security Institute (AISI) published “Towards technical best practice for frontier AI development” on 12 October, the same day as the AI minister's speech. It sets out what the institute thinks frontier AI developers should standardise next.

It names two challenges. “First, risks are arising earlier in the development process”: today's systems carry risk while unreleased, “during training, internal deployment, and evaluation itself.” Second, evaluation is struggling to keep up, as benchmarks saturate quickly and alignment and control methods are still immature.

The post lists eight priority areas. In outline they are: assessing and documenting risk; overseeing AI systems through their lifecycle; preserving and evaluating oversight mechanisms; detecting, responding to and reporting incidents involving AI agents; sharing information on how fast AI is developing; enabling independent auditors and evaluators; measuring capabilities in security-relevant fields; and measuring alignment. On incidents, it asks how unsafe agent behaviour should be defined and its severity judged, and how incidents should be reported to affected parties, governments and the public.

The institute says it “is not seeking to mandate requirements for frontier developers”. It will work with developers on voluntary practices, which it says could be formalised through international standards organisations. It rests its standing on experience, saying that, working with industry, “we have independently evaluated more than 55 models before release”. No authors are named, and there is no separate full document.

• Confirmed AISI's 12 October post sets out eight priority areas for technical best practice, including detecting, responding to and reporting agentic AI incidents. AI Security Institute
• Confirmed AISI says it “is not seeking to mandate requirements for frontier developers”; practices could be formalised via international standards development organisations. AI Security Institute
• Claimed AISI says it has independently evaluated more than 55 models before release. AI Security Institute