Library: last editorial review 20 Sept 2026. Check the original source for current status; the automated radar is updated separately.
Library
Documents, rules and research on AI policy. Filter by theme, scope or source; each reference retains its review date.
The Commission’s overview of GPAI obligations, including documentation, copyright and training-content transparency, with additional evaluation, risk-management, incident-reporting and cybersecurity duties for systemic-risk models. Exceptions and transition periods must be checked in the regulation.
The 2026 international scientific synthesis on general-purpose AI capabilities, misuse, loss of control and risk management. Read its evidence, uncertainties and disagreements separately from policy choices; it is not a certification of a particular model.
NIST’s November 2024 announcement of the International Network of AI Safety Institutes. This is a historical founding document, not a current membership register; it cannot establish whether a country belongs to the network in September 2026.
The 2023 international declaration, with Spain listed among represented countries, that frames frontier AI risks as international and calls for shared science, evaluation, transparency, and public-sector capability.
A practical risk-management bridge for public and private organizations: govern, map, measure, and manage AI risks, with a 2024 generative-AI profile and 2026 critical-infrastructure work in progress.
Survey of AI researchers conducted in 2023, reporting respondents’ beliefs about progress and risks. Responses depend on the question wording and sample; they are elicited opinions, not observed frequencies or calibrated probabilities of future catastrophe.
METR’s March 2025 paper on task-completion horizons, reporting an approximately seven-month doubling trend over its historical sample. The result concerns the evaluated tasks and reliability threshold; it is not a forecast or proof of general job automation.
METR’s live task-horizon dashboard, showing 50% and 80% success thresholds. A horizon refers to task difficulty measured by human completion time, not how long an agent reliably works alone. The page’s displayed update is 8 May 2026; the current suite has limitations for tasks above 16 hours.
A policy-useful map of frontier capability benchmarks across reasoning, coding, agents, multimodality, and other domains, useful for avoiding single-score model hype.
Epoch AI's expert-level mathematics benchmark, designed to test advanced reasoning beyond saturated standard exams and to support more serious capability measurement.
The Nature paper for Humanity's Last Exam, a 2,500-question expert benchmark across subjects; useful because it shows both harder evaluation design and continuing uncertainty.
The Terminal-Bench contributors’ historical 2.0 release, published on 7 November 2025, for agentic work in terminal environments. The primary project lists subsequent versions, so 2.0 must not be treated as the latest release or compared across versions without qualification.
A benchmark for multimodal agents doing open-ended tasks in real computer environments; relevant to public-sector AI because autonomy increasingly means operating software, not just generating text.
A concrete evaluation toolkit for dangerous autonomous capabilities, including task-suite guidance and an example protocol for measuring agentic frontier-model risk.
OpenAI’s October 2023 announcement of its approach to frontier risk. Retained as historical primary-source material, not as the latest version of its framework or an independent assessment of the company’s models.
OpenAI’s Frontier Governance Framework, announced on 28 May 2026. It describes the company’s approach to severe risks, safety and security practices, reporting and emerging legal requirements. A company framework is not an independent audit or a finding of compliance.
Anthropic’s February 2026 announcement of RSP v3.0 and its Frontier Safety Roadmap approach. This is a dated statement of the company’s policies and goals, not independent evidence of model safety or a guarantee that the policy remains unchanged.
Google DeepMind’s dated announcement on its Frontier Safety Framework. It describes capability thresholds and risk-management practices proposed by the laboratory; it should not be read as an independent evaluation or a certification of compliance.
The 2023 Statement on AI Risk, recording its signatories’ concern about extreme AI risks. This is an advocacy statement, not a scientific estimate of probability, a legal rule or evidence that a particular model poses that level of risk.