On 12 October 2026 the UK AI Security Institute (AISI) published an article setting out eight priorities for technical practice in frontier AI safety. Its central judgment is that, as frontier AI capabilities advance rapidly, safety risk is moving forward into the model development stage, and that existing methods of safety assessment struggle to keep pace with technical development, so new technical safety practices need to be explored. This echoes another development a few days earlier: on 11 October AISI resumed pre-release access to and testing of new models from Anthropic, OpenAI and Google DeepMind.
The content of the eight practices
The eight priorities cover the full chain from development to deployment, and from capability evaluation to incident response:
First, risk assessment and recording throughout development. Risk assessment is to be embedded in the whole of model development and recorded, rather than carried out once before release. This runs in the same direction as the security assessment requirements of China’s Interim Measures for the Administration of Generative AI Services; the difference is that AISI emphasises the whole process and traceability.
Second, supervision across the AI system life cycle. Supervision extends from training to deployment and operation.
Third, maintaining and evaluating the oversight mechanism. The oversight mechanism itself must be maintained and periodically evaluated, so that it does not lapse once established.
Fourth, incident response and reporting for AI agent safety. For agents — forms able to call tools autonomously and interact with external systems — a dedicated process of safety incident response and reporting should be established.
Fifth, sharing information on the pace of AI development. Information on the rate of capability development should be shared between industry and regulators, so that assessment methods do not fall behind technical evolution.
Sixth, independent review and evaluation. Third parties independent of the developer should be brought into review and evaluation.
Seventh, evaluation of capability in safety-relevant domains. Model capability in safety-relevant domains should be assessed specifically.
Eighth, evaluation of AI alignment. Model performance on value alignment should be assessed.
Why assessment should move forward
The common direction of the eight practices is to shift the centre of gravity of safety work from a single gate before release to management across the whole of development. AISI puts it directly: existing methods of safety assessment struggle to keep pace with technical development. The problem, in other words, is not that assessment is insufficiently strict, but that its timing and method cannot keep up with the rate at which model capability iterates.
For developers this means two things. First, the outputs of safety assessment — assessment records, risk inventories, alignment test results — are themselves verifiable material, and must be accumulated from the early stages of a project rather than assembled at the point of release. Second, assessment ceases to be a single act and becomes a continuing process that must be maintained and independently reviewed.
Experience from pre-release evaluation of 55 models
The article discloses one figure: AISI has worked with the AI industry to conduct independent evaluation of more than 55 AI models before their public release, and has used that work to develop principles for safety safeguard assessment. The number shows that independent pre-release evaluation in the UK has moved from an ad hoc arrangement to a routine mechanism.
What is notable is AISI’s next step: it states that, as these practices mature, they will need to be converted into formal technical standards through the relevant international standards bodies. This is an upgrade path from a regulator’s technical advice to an international standard. For companies going overseas, it means these practices may in future enter compliance requirements in the form of standards rather than guidance alone.
Relationship with the ICO’s regulatory path
The UK currently has two parallel lines of AI regulation. One is the Information Commissioner’s Office approaching from data protection, first subjecting foundation model developers to continuing supervision and then extending regulation to AI agents. The other is AISI approaching from technical safety, focusing on the development process and pre-release assessment of frontier models. The two lines have overlapping objects but different concerns: the ICO is concerned with personal data and the rights of data subjects, AISI with the safety risk arising from model capability.
The practical consequence for companies is that, addressing the UK market, two sets of material must be prepared at the same time — transparency and rights-exercise mechanisms on the data protection side, and risk assessment, alignment evaluation and incident response records on the safety side. On the ICO’s requirements, see The UK ICO’s Regulatory Progress on Foundation Models and AI Agents.
What it means for Chinese companies going overseas
These practices offer Chinese AI companies two points of direct reference.
First, make assessment records a routine output. Several of the eight practices — risk assessment and recording throughout development, independent review and evaluation, alignment evaluation — are in substance requirements to keep a record. Test records, risk assessment documents and descriptions of guardrail design are engineering management beforehand and compliance assets afterwards.
Second, take agent-related safety incident response seriously. The fourth practice is addressed specifically to AI agents, and is concerned with the same category of risk as the “unexpected autonomous behaviour” and “autonomous cyber-attack threat” newly added in China’s AI Safety Governance Framework 3.0. Making controllability, traceability and accountability robust is one of the few directions that satisfies both the Chinese and UK regulatory requirements at once.
For assistance with preparing data protection and safety assessment materials for overseas products, and with multi-jurisdiction compliance arrangements, see the EU AI Act and Multi-Jurisdiction Compliance service pages. Related reading: the analysis of the AI Safety Governance Framework 3.0 discusses domestic requirements on the autonomous behaviour of agents, and AI compliance for overseas markets under dual-track regulation sets out the parallel arrangements of the two sets of rules.
This article is prepared on the basis of the article on technical practices for frontier AI safety published by the UK AI Security Institute on 12 October 2026, and a report by Cailian Press on 13 October 2026. It is for reference only and does not constitute legal advice. Specific compliance judgments should be made in the light of the applicable jurisdiction and the facts.