Catastrophic Risk

The operational EU entry point for systemic-risk GPAI: all GPAI providers face documentation, copyright, and training-content duties; systemic-risk models add notification, risk mitigation, incident reporting, and cybersecurity obligations.

The best current scientific baseline for advanced general-purpose AI risk: capabilities, misuse, loss-of-control debates, risk-management techniques, and evidence gaps, backed by 30+ countries and international organizations.

The frontier-model evaluation network to compare with Spain's architecture: Spain participates in international safety-report processes and has AESIA for AI Act supervision, but is not publicly listed as a national AI Safety Institute member.

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.

A survey of 2,778 AI researchers: the wiki reports 5%, 10%, and 5% median responses across three extinction or severe-disempowerment phrasings, plus 57.8% giving at least 5% odds to extremely bad HLMI impacts.

METR's task-horizon work makes autonomy legible for policy: the March 2025 paper reports roughly seven-month doubling over six years, while the newer live estimates are faster and should be read separately.

METR's live autonomy dashboard: 50% and 80% reliability time horizons for public frontier agents. The 8 May 2026 update adds Claude Mythos Preview (early), about a 17.4-hour p50 and 3.1-hour p80 estimate, plus an explicit warning that measurements above 16 hours are unreliable with the current suite.

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.

Epoch AI's terminal-task benchmark for agentic command-line work, a useful proxy for software operations, debugging, and tool-using capability.

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.

A frontier-lab primary source on severe-risk categories, including cybersecurity, CBRN, persuasion, autonomy, evaluations, red-teaming, reporting, and security controls.

A 28 May 2026 frontier-lab governance source mapping safety and security practices to emerging legal requirements, including the EU AI Act GPAI Code of Practice track, cyber, CBRN, manipulation, loss-of-control, reporting, security, and incident response.

A frontier-lab primary source on Anthropic's RSP v3.0: it replaces the earlier hard-pause style trigger with public Frontier Safety Roadmaps, goals, risk reports, safeguards, evaluations, and security requirements.

DeepMind's updated frontier-safety framework adds tracked capability levels and a risk-management process for severe risks from advanced models, useful for comparing lab governance with public oversight.

An influential 2023 public statement that helped push extinction-risk language into mainstream debate; include it as a signal of expert concern, not as a substitute for evidence, law, or institutional design.