International AI Safety Report 2025 published ahead of Paris AI Action Summit

The International AI Safety Report 2025 is a government-commissioned scientific synthesis of research on the capabilities and risks of general-purpose AI, chaired by Yoshua Bengio. It was published on 29 January 2025, ahead of the AI Action Summit held in Paris the following month.

What it is

The report is described as the work of 96 international AI experts who collaborated to establish a shared scientific understanding of risks from advanced AI and methods for managing them. It builds on an Interim Report published in May 2024, which was presented at the AI Seoul Summit. The effort followed the commitment made by countries at the Bletchley Park AI Safety Summit to support creation of the report. The work was overseen by an international Expert Advisory Panel nominated by over 30 countries and intergovernmental organisations.

What it requires

The report is not a binding legal instrument but a synthesis of existing research intended to inform policymakers, and it does not represent the views of any government that supported its development. It groups risks from advanced AI into malicious use, technical failures and systemic risks, covering issues such as large-scale scams, cyberattacks and sexualised deepfakes. On technical failures, it warns that AI systems may malfunction or behave unpredictably due to unreliable reasoning, poor generalisation, or mis-specified objectives. On systemic risks, it points to increased dependence on a small number of model providers and the potential for cascading failures across interconnected infrastructures.

What security leaders should do

Security leaders should treat the report as a reference document for board and risk-committee briefings on general-purpose AI capability growth, since it tracks evidence such as models showing markedly better performance at tests of programming and scientific reasoning. They should also factor the report's observation that many companies are investing in general-purpose AI agents that can autonomously plan and act with little or no human oversight into agentic-AI risk assessments. Because the report notes that sophisticated attackers can often bypass current defences and the real-world effectiveness of many safeguards is uncertain, teams should avoid treating any single vendor safety framework as sufficient. Security leaders should also monitor successor editions, since a second full International AI Safety Report was released on 3 February 2026 ahead of a subsequent summit.