International AI Safety Report 2025

Summary

The first full International AI Safety Report, the work of 96 AI experts with an Expert Advisory Panel nominated by 30 countries, the OECD, the EU and the UN. It summarises the scientific evidence on the capabilities of general-purpose AI, including AI agents; on its risks from malicious use (fake content, manipulation, cyber offence, biological and chemical attacks), malfunctions (reliability, bias, loss of control) and systemic effects (labour markets, the AI R&D divide, market concentration, environment, privacy, copyright) and open-weight models; and on technical approaches to identifying, assessing, mitigating and monitoring risks. It does not recommend specific policies.

Safety topics

Status notes

Published ahead of the AI Action Summit in Paris in February 2025, following the interim report of May 2024 (INT-08), which was presented at the AI Seoul Summit. A Chair's note covers AI advances between the end of the writing period (5 December 2024) and publication, notably early test results of OpenAI's o3 model. The UK Government said it would keep providing the secretariat until a long-term international home was agreed, with Yoshua Bengio as chair for 2025. Two Key Updates followed as separate publications: 'First Key Update: Capabilities and Risk Implications' (15 October 2025) and 'Second Key Update: Technical Safeguards and Risk Management' (25 November 2025). The second edition, the International AI Safety Report 2026 (INT-10), followed on 3 February 2026. The PDF's back page prints the research series number as DSIT 2024/000; its front matter and the website give DSIT 2025/001.

Sources

Open the official source https://internationalaisafetyreport.org/publication/international-ai-safety-report-2025

Secondary source: https://www.gov.uk/government/news/first-international-ai-safety-report-to-inform-discussions-at-ai-action-summit

Local copy

Verification

History

1. 2026-10-18related: [{"slug": "international-scientific-report-safety-advanced-ai-interim", "id": "I → [{"slug": "international-scientific-report-safety-advanced-ai-interim", "id": "I — two-way link added: INT-05, SG-07 link(s) to INT-092026-10-18-fix-related-links · Claude Code (claude-opus-5-5)
2. 2026-10-17notes: Published for the Paris AI Action Summit. Two interim 'Key Updates' followed: 'F → Published ahead of the AI Action Summit in Paris in February 2025, following the — rewritten: the Paris summit and the interim report from Bengio's foreword, the Chair's note on o3, the secretariat arrangement from the GOV.UK release, the Key Updates and the 2026 edition with their dates from the Publications page stored in this run, and the DSIT number misprint on the PDF's back page; the unsupported GOV.UK redirect sentence removed2026-10-17-recheck-international-2 · Claude Code (claude-opus-5-5)
3. 2026-10-17summary: The first full international scientific assessment of general-purpose AI capabil → The first full International AI Safety Report, the work of 96 AI experts with an — checked against 'About this report' and the contents: 96 experts (not 'about 100'), the 30 countries, OECD, EU and UN, the risk sub-sections and the risk-management chapter listed; 'It does not recommend specific policies' kept as the report words it2026-10-17-recheck-international-2 · Claude Code (claude-opus-5-5)
4. 2026-10-17date_last_updated: 2025-11-25 → – — 2025-11-25 was the date of the Second Key Update, a separate publication between editions, not an amendment or new version of the 2025 report; the Key Updates now go in the notes2026-10-17-recheck-international-2 · Claude Code (claude-opus-5-5)
5. 2026-10-17issuing_body: Independent expert panel chaired by Yoshua Bengio; secretariat in the UK AI Secu → Independent expert panel chaired by Yoshua Bengio, with a secretariat at the UK — the stored contributor lists name two secretariat bodies, the AI Safety Institute and Mila – Quebec AI Institute; the old value named only the UK institute2026-10-17-recheck-international-2 · Claude Code (claude-opus-5-5)
6. 2026-10-14language: English (translations published) → English — The field held a remark after the language name; it now gives the language names only (AGENTS.md 5.6).manual-2026-10-14 · Claude Code (claude-opus-5-5)

Related records

International AI Safety Report (interim, 2024) International AI Safety Report (AI Safety Summit process) · Superseded

International AI Safety Report 2026 International AI Safety Report (AI Safety Summit process) · Published (non-binding, current)

Paris AI Action Summit Statement AI Action Summit (Paris, 2025) · Published (non-binding, current)

Seoul Statement of Intent on AI Safety Science AI Seoul Summit (2024) · Published (non-binding, current)

Singapore Consensus Singapore · Published (non-binding, current)

Citation

Independent expert panel chaired by Yoshua Bengio, with a secretariat at the UK AI Safety Institute and Mila – Quebec AI Institute; published by the UK Government (DSIT). International AI Safety Report 2025 [DSIT 2025/001]. 2025-01-29. https://internationalaisafetyreport.org/publication/international-ai-safety-report-2025. Status: Superseded; verified 2026-10-17 (level A). Local copy archive/INT-09/2026-10-11/official.html retrieved 2026-10-11 13:57:29 UTC, SHA-256 0fa3a9e56890013e95da8bb888d329c25a2554ed809e789b683af30f7067012b. Every AI Policy Ever, version 1.18 (snapshot 2026-10-18), record INT-09.

Improve this record

Know of an amendment, a status change or an error? Report news or an error on GitHub; the form opens with this record filled in.