Standards and Regulation: UK AI Security Institute

中文版

Standards and Regulation · UK AI Security Institute. Frontier model evaluations, the Inspect framework and the international report secretariat. This source has 5 entries admitted under this category's review policy; each links directly to the original.

Authority tiers in this category: A1.

• Inspect: the UK AISI's open-source model evaluation framework

  Publisher: UK AI Security Institute · Collection: AISI Evaluation Tools · Format: Official evaluation framework · Authority tier: A1 · Review status: Passed single-tier standards and regulation review

  Provides a reproducible system of tasks, datasets, solvers, scorers and logs for building and running model evaluations.

  Evidence use: Supports Inspect's official status, components and evaluation uses.

  Limitations: Tool results depend on the task and configuration. The official release note uses a blog URL, but the subject itself is a formal open-source framework.

  Related concepts: Model Evaluation and Benchmarks · AI Red Teaming

  Link last checked: 2026-09-24

• UK AISI model evaluation approach

  Publisher: UK AI Security Institute · Collection: AISI Evaluation Guidance · Format: Official evaluation methodology · Authority tier: A1 · Review status: Passed single-tier standards and regulation review

  Explains AISI's goals for independent technical evaluation, risk areas, methodological limits and its position of not issuing "safety certification".

  Evidence use: Supports AISI's evaluation role, scope and the limits of its conclusions.

  Limitations: AISI states explicitly that it is not a regulator. Preliminary evaluations do not constitute safety claims.

  Related concepts: Model Evaluation and Benchmarks · AI Red Teaming

  Link last checked: 2026-09-24

• UK AISI research agenda

  Publisher: UK AI Security Institute · Collection: AISI Research Agenda · Format: Official research and evaluation framework · Authority tier: A1 · Review status: Passed single-tier standards and regulation review

  Defines priority risks and methods in cyber, chemical and biological, autonomous systems, human influence, the science of evaluations and safeguards analysis.

  Evidence use: Supports AISI's current research and evaluation priorities.

  Limitations: A research agenda for a given period, not a regulatory requirement. Sensitive methods are not fully disclosed.

  Related concepts: Model Evaluation and Benchmarks · AI Red Teaming · AI Governance and Regulation

  Link last checked: 2026-09-24

• Principles for frontier AI misuse safeguard evaluations

  Publisher: UK AI Security Institute · Collection: AISI Safeguard Evaluation · Format: Evaluation principles and templates · Authority tier: A1 · Review status: Passed single-tier standards and regulation review

  Sets out core principles for evaluating misuse safeguards, with an accompanying template of actionable questions.

  Evidence use: Supports the recommended principles and question structure for assurance evaluations.

  Limitations: A recommended practice, not a mandatory rule. Methods will be updated as frontier capabilities develop.

  Related concepts: AI Guardrails · Model Evaluation and Benchmarks · AI Red Teaming

  Link last checked: 2026-09-24

• Frontier AI Trends Report

  Publisher: UK AI Security Institute · Collection: AISI Public Assessments · Format: Official evidence assessment report · Authority tier: A1 · Review status: Passed single-tier standards and regulation review

  Summarizes AISI's public evidence assessment of frontier model capabilities, risks and trends in safeguards.

  Evidence use: Supports frontier trend assessments and AISI measurement results as of the report date.

  Limitations: A point-in-time assessment, not a regulatory rule. Conclusions about fast-changing models must be tied to a version.

  Related concepts: Model Evaluation and Benchmarks · Scaling Laws · AI Governance and Regulation

  Link last checked: 2026-09-24