First International AI Safety Report published

Background

"Frontier AI" refers to the most capable general-purpose systems available at a given moment — models trained on broad data that can be applied to many different tasks rather than one narrow job. Because a single system of this kind can be pointed almost anywhere, no one country's testing tells the whole story. The report was commissioned following the 2023 Bletchley AI Safety Summit, where 30 nations agreed to build a shared scientific baseline on frontier AI risks and capabilities^[2]. A baseline in this sense is not a minimum safety standard that countries commit to; it is a common description of the evidence, so that governments drafting rules start from the same account of what is known rather than from separate ones.

What was published

The first International AI Safety Report was published on 29 January 2025^[1]. It is led by Yoshua Bengio, a Turing Award winner, and was authored by more than 100 AI experts; 30 countries and international organisations back the project^[1]. The expert advisory panel was nominated by the 30 participating countries together with the UN, the EU and the OECD^[1][2]. Its members include Yi Zeng of the Chinese Academy of Sciences and Saif M. Khan of the US Department of Commerce^[1].

The work was commissioned by the UK Government and supported by a Secretariat within the UK AI Security Institute^[2]. It also did not arrive in one step: an Interim Report was published in May 2024 in advance of the AI Seoul Summit, and the first annual report came ahead of the France AI Action Summit in 2025^[2].

The three questions

The report is organised around what general-purpose AI systems can do, what risks they pose, and what techniques exist to mitigate those risks^[2]. Capability, in this framing, means what a system can actually be used for and how reliably; risk covers the ways that use can go wrong; and mitigation means the technical measures that make a harm less likely or less severe — for instance methods that make a system's behaviour easier to check — as distinct from laws or rules imposed on developers. Keeping to technical description is deliberate: the report provides scientific information for policymaking but makes no policy recommendations^[2]. It sets out what is known and where the evidence is thin, and leaves the choice of what to regulate, and how, to the governments and other bodies using it.

Aftermath

Bengio will continue as chair through 2026^[2]. Because the 2025 document is presented as the first annual report rather than a one-off, the shape it establishes — a recurring, multi-country reference on the evidence, with panels nominated by participating governments and by the UN, EU and OECD^[1][2] — is meant to be repeated. Its usefulness as a reference rests on the breadth of that panel: the membership includes researchers and officials from both China and the United States^[1], countries whose regulatory approaches differ, which is part of the point of a document that describes the science without recommending policy^[2].