Second “Key Update” of the International AI Safety Report Released

The latest update to the International AI Safety Report has been published, marking an important milestone in the global effort to govern advanced AI. Chaired by Turing Award-winner Yoshua Bengio, the report draws on contributions from more than 100 international experts and is backed by over 30 countries and major institutions including the European Union, the Organisation for Economic Co‑operation and Development (OECD) and the United Nations. 

Originally introduced as an annual publication, the report now issues shorter “Key Updates” to provide timely guidance on rapidly-evolving AI risks and mitigation strategies. This second update follows its first edition in October 2025. 

Key Findings

• Defensive measures for AI are improving, but “roughly 50%” of sophisticated attacks still succeed with just ten attempts. Training-data poisoning remains a persistent threat with as few as 250 malicious documents needed to compromise a model. 
• The gap between research and open-source model releases is narrowing: leading open-weight models are now less than a year behind closed-source leaders, increasing accessibility but also raising misuse concerns. 
• Although the number of AI firms adopting risk-management frameworks at least doubled in 2025, there is limited evidence yet that they are consistently effective in practice. 

Why It Matters

With national regulators preparing for more far-reaching AI policies, this update serves as a timely roadmap. It emphasizes both technical and institutional gaps — from cyber-threat resilience to governance practices — and encourages adoption of evidence-based standards. Decision-makers at governments, multilateral institutions and industry will find in the report a consolidated view of where the field stands and what needs attention ahead of the full report expected in early 2026.