International AI Safety Report 2026

What it claims

This is a synthesis of existing research, not new results, and it makes no specific policy recommendations. Its scope is narrower than the 2025 edition. It covers only "emerging risks" at the frontier of general-purpose AI. It drops the 2025 edition's coverage of bias, environmental impacts, privacy and copyright, and leaves those to efforts such as the UN's Independent International Scientific Panel on AI. New in this edition are scenarios and forecasts drawn from research by the OECD and the Forecasting Research Institute. The page names over 30 nominating countries and organisations, which is more than the 29 countries in the existing summary.

How to use it

The report is licensed under the Open Government Licence v3.0 "except where otherwise stated", © Crown 2026, research series DSIT 2026/001. Material can generally be reused under OGL terms, but check individual figures for other licences. Citation guidance is on p. 156.

Sections most relevant to engineers, by page:

• 1.2 Current capabilities (p. 26) and 1.3 Capabilities by 2030 (p. 32)
• 2.1.3 Cyberattacks (p. 57) and 2.1.4 Biological and chemical risks (p. 64)
• 2.2.1 Reliability challenges (p. 71) and 2.2.2 Loss of control (p. 76)
• 3.2 Risk management practices (p. 105). The Indian minister's foreword says it covers model evaluations, dangerous-capability thresholds and "if-then" safety commitments.
• 3.3 Technical safeguards and monitoring (p. 120)
• 3.4 Open-weight models (p. 132)

Other risk sections cover AI-generated content and crime, influence and manipulation, labour markets and human autonomy.

Named contributors include Yoshua Bengio (Chair) and writers from Epoch AI, Apollo Research, the Centre for the Governance of AI (GovAI) and others. Industry reviewers included Anthropic, OpenAI, Google DeepMind, Meta, Microsoft, Hugging Face and others, alongside civil-society organisations.

Limits

• No findings were fetched. The fetched text covers only the front matter: contributors, forewords, contents and the start of "About this Report". None of the report's evidence, numbers or forecasts appear here, so the actual content needs the full PDF.
• No endorsement of the conclusions. The report states it does not necessarily represent the views of the Chair, individual contributors or supporting governments. The Expert Advisory Panel gave technical feedback only.
• Immature evidence base. The Indian government foreword describes risk management frameworks as "still immature, with limited quantitative benchmarks and significant evidence gaps."
• Missing topics. Readers who need bias, privacy, copyright or environmental coverage should use the 2025 edition.