--- title: "AISI Names Eight Technical Priorities for Frontier Labs, and Asks Nicely" pubDatetime: 2026-10-20T21:53:00.000Z description: "The UK AI Security Institute says frontier AI risk now emerges during training and internal deployment, not at launch, and that evaluations are falling behind. Its answer is eight voluntary priority areas it has no power to require." tags: [aisi, ai-policy, evaluations, ai-security, ai-agents, standards, frontier-labs, 2026, 2026-q4, 2026-10] --- > [!tldr] TL;DR > On 12 October the UK AI Security Institute published eight priority areas where it wants new technical best practice for frontier AI development, built with industry rather than imposed on it. The reasoning behind the list is the interesting part: AISI argues risk now appears during training, internal deployment and evaluation, well before a model ships, and that evaluation methods are falling behind capability growth. AISI says it is "not seeking to mandate requirements for frontier developers, but to contribute our expertise to an industry-wide effort," with formalisation through international standards bodies left for later. It cites more than 55 models independently evaluated before release as the basis for its expertise. No source gives a date or milestone for the voluntary-to-standard transition. The UK AI Security Institute evaluates frontier models before they ship, and it has now said in public that the risks it looks for have moved upstream of that job. Release day used to be where the risk concentrated. In [its 12 October post](https://www.aisi.gov.uk/blog/towards-technical-best-practice-for-frontier-ai-development), AISI puts it earlier: during training, during internal deployment, and during evaluation itself. That reframing matters more than the list it supports. A pre-release evaluation regime assumes a clean boundary between "model under development" and "model in the world." Internal deployment erases the boundary. A lab running a cyber-capable agent inside its own infrastructure has created exposure that no release-day sign-off touches. ## Eight priorities, one of them visible in the coverage AISI sets out eight areas it considers most urgent and calls the list non-exhaustive. Trade coverage the following day, including [On The Wire's summary](https://onthewire.ai/article/uk-ai-security-institute-sets-out-eight-priorities-it-says-frontier-ai-developer), confirms the count and picks out one concrete item: reporting incidents caused by AI agents. Agent incident reporting is a good test of whether any of this works. It asks a lab to publish a record of its own system doing something it should not have done, in a period where the lab is under no obligation to publish anything. Everything else on an eight-item list of good practice can be adopted quietly. This one is visible or it has not happened. ## The evaluator says the evaluations are slipping AISI's second core claim is that evaluation methods are struggling to keep pace with capability growth. Benchmarks saturate quickly. Credible multi-agent evaluations take longer to build than the systems they are meant to measure. Read that as self-assessment rather than a neutral survey of the field. AISI has run more than 55 independent pre-release model evaluations and open-sourced the tooling it uses, Inspect. The body with the best view of whether frontier evaluation is keeping up has said it is not, which is a sharper admission than the tidy "priorities" framing suggests. A list of areas needing work is also a list of places where the current method falls short. ## Voluntary by design, and that is the weak point AISI wants none of this mandated. The path it describes runs from voluntary practice, developed with industry, toward eventual formalisation through international standards bodies. That is soft power by choice, and it fits an institute whose leverage comes from technical credibility rather than statutory authority. The objection writes itself. Frontier labs combine the largest capability advantage with the thinnest external enforcement, so a voluntary framework asks the party with the most to lose from oversight to supply it. AISI's own diagnosis sharpens the problem. If risk emerges during training and internal deployment, the practices that matter govern work no outside party observes. Voluntary incident reporting is worth exactly as much as a lab's willingness to file an embarrassing report. The counter-case is that mandates need something to mandate. Nobody can write a regulation requiring good multi-agent evaluation practice while the field cannot agree what that looks like. Standards bodies move on multi-year cycles and need technical consensus as input. On that reading the voluntary phase is the drafting phase, and AISI's bet is that convening labs now yields a better standard in three years than legislating badly today. Timing is where the bet gets tested, and it is where the post goes quiet. No source establishes a deadline or milestone for moving from voluntary practice to international standard. An iterative path with no checkpoint is hard to tell apart from an indefinite one, and "risk emerges earlier" is itself an argument that there is less time than a standards track allows. ## What would count as follow-through Two things would make this concrete. First, a frontier lab publicly adopting the priorities and saying which ones it already meets. Second, AISI publishing the practices rather than the areas, since a priority area is a heading and a best practice is a document someone can be held to. Eight days after publication, the available coverage shows neither. What does AISI do if the same is true a year from now? --- *Sources: [Towards technical best practice for frontier AI development](https://www.aisi.gov.uk/blog/towards-technical-best-practice-for-frontier-ai-development)* *Coverage: [UK AI Security Institute sets out eight priorities](https://onthewire.ai/article/uk-ai-security-institute-sets-out-eight-priorities-it-says-frontier-ai-developer) • [UK AISI proposes eight priorities for frontier AI best practice](https://aidaily.si/stories/uk-aisi-proposes-eight-priorities-for-frontier-ai-best-practice.html)* *Related on this blog: [Agent Harnesses: A Standard for a Word Nobody Agrees On](/posts/agent-harnesses-standard) • [GPT-6 Astra Ran Supply-Chain Attacks in 29% of AISI's Simulated Trajectories](/posts/aisi-gpt-6-astra-supply-chain-attacks) • [[AUTO] AISI finds AI agents coordinating to inject malware into open-source](/posts/auto-agents-malware-injection)*