MONTREAL, Nov. 25, 2025 /PRNewswire/ -- The second Key Update of the International AI Safety Report has been published, providing a timely update on risk management and technical mitigations for general-purpose AI.

Chaired by Turing Award-winning computer scientist Yoshua Bengio, the Report involves more than 100 international experts and is backed by over 30 countries and international organisations including the EU, OECD and UN.

Recognising that the field is evolving too quickly for a single annual Report, Key Updates were introduced to provide shorter, focused reports on critical developments. This Update follows the first Key Update, which was published on October 15, 2025. These updates ensure policymakers have an updated synthesis of the literature to inform evidence-based policymaking.

The second Key Update covers an overview of technical methods for improving reliability and preventing misuse, as well as emerging institutional approaches to risk management. Its key highlights include:

• Progress is being made to train models to resist malicious attacks, but gaps remain. While AI models and systems have become better at resisting attacks, sophisticated hackers can still break safeguards in roughly 50% of cases with just 10 attempts, and compromise models by poisoning their training data with as few as 250 malicious documents.
• The open-source gap is closing. Open-weight models now are less than a year behind industry leaders, democratising access but complicating efforts to prevent malfunctions and misuse.
• Industry safety commitments have increased but effectiveness is unknown. Although the number of AI companies adopting risk management frameworks more than doubled in 2025, their effectiveness in practice remains uncertain.

Chair of the Report Yoshua Bengio, Full Professor at Université de Montréal, Scientific Director of LawZero and Scientific Advisor of Mila - Quebec AI Institute, said:

"As we continue to follow updates on AI's capabilities and risks, it is essential to also provide clear pathways for appropriate and effective risk management and technical mitigations. This Key Update provides an overview of the progress being made in these areas, as well as the gaps and opportunities which still remain. Our aim is to continually provide decision-makers around the world with timely updates on AI's developments, in advance of the second International AI Safety Report which is set to be published in early 2026, ahead of the AI Impact Summit in India."

About

The International AI Safety Report is a synthesis of the evidence on the capabilities and risks of advanced AI systems. It is designed to support informed policymaking globally by providing an evidence base for decision-makers. Authored by a diverse group of over 100 independent experts, the report is backed by an Expert Advisory Panel composed of nominated representatives from over 30 countries and international organisations including the EU, OECD and UN. While acknowledging AI's immense potential benefits, the Report's focus is on identifying risks and evaluating mitigation strategies to ensure AI is developed and used safely for the benefit of all. The Report was commissioned by the UK Government with the Secretariat sitting in the UK AI Security Institute.

Contact: Mila Medias (medias@mila.quebec); DSIT Media Enquiries (press@dsit.gov.uk)