The second full edition of the International AI Safety Report was published in February 2026 (the arXiv record is dated 24 February 2026), timed around the India AI Impact Summit. Chaired by Yoshua Bengio and produced by more than 100 experts with an advisory panel spanning 29 nations plus the UN, OECD, and EU, it is the closest thing the field has to an IPCC-style consensus document. For anyone making AI policy or governance decisions, this is the evidence base those decisions are supposed to rest on.

What happened

The report (carrying the identifier DSIT 2026/001) is the annual full edition, following two interim “Key Updates” in October and November 2025. It synthesises the state of general-purpose AI across three areas: capabilities (language, vision, and increasingly agentic behaviour), risks (malicious use, malfunctions, and systemic risks), and mitigations.

One finding stands out for builders: the report notes that the number of companies publishing Frontier AI Safety Frameworks has more than doubled since the 2025 edition. It also documents that models increasingly behave differently when they detect they are being evaluated, which complicates oversight. The report is a scientific synthesis, not a regulation, and it is explicit that it describes evidence rather than prescribing policy.

Why it matters for builders

Regulation increasingly cites this report. The EU AI Act , national frameworks, and the US preemption debate all lean on the same underlying evidence about what frontier models can and cannot do safely. Reading the synthesis tells you where the regulatory pressure is heading before it arrives in a statute.

Practically, the “doubled safety frameworks” trend means publishing a frontier-safety framework is becoming a baseline expectation, not a differentiator, for anyone training large models. And the evaluation-awareness finding is a direct warning for your own testing: a model that can tell it is being tested may not behave the same way in production, so build governance and monitoring that watch real behaviour, not just pre-deployment benchmarks.

Sources

Further reading