#Canada #Montreal #AI Safety #Yoshua Bengio #International Report

Second Key Update of the International AI Safety Report Released

On November 25, 2025, the second Key Update of the International AI Safety Report was made public, presenting critical insights into the management of risks and technical mitigations associated with general-purpose artificial intelligence (AI). Chaired by renowned computer scientist and Turing Award winner Yoshua Bengio, this report incorporates contributions from over 100 international specialists and garners backing from more than 30 countries and global organizations, including the European Union (EU), the Organisation for Economic Co-operation and Development (OECD), and the United Nations (UN).

With the rapid evolution of AI technology rendering annual reports insufficient, the introduction of Key Updates aims to offer concise assessments on major advancements. This latest update follows the first Key Update that was released on October 15, 2025, ensuring that policymakers have access to updated information to facilitate evidence-based decisions regarding AI.

Key Highlights of the Update

The second Key Update elaborates on various technical strategies for enhancing system reliability and mitigating potential misuse of AI technology. Here are some critical highlights from the report:

1. Progress on Model Resilience: Researchers are making strides in training AI models to withstand malicious attacks. However, vulnerabilities still persist. Despite improvements, sophisticated cyber attackers can manipulate protective systems nearly half the time after only ten attempts. Notably, models can be compromised using as few as 250 harmful training documents.

2. Shrinking Open Source Gap: The disparity between open-source AI models and industry frontrunners is diminishing significantly, with open-weight models now trailing by less than a year. This democratization of access is a double-edged sword, complicating efforts to prevent malfunctions and misuse.

3. Increased Industry Commitment to Safety: The number of AI enterprises implementing risk management frameworks doubled in 2025. However, the actual effectiveness of these frameworks remains uncertain.

Yoshua Bengio, who serves as the Full Professor at the Université de Montréal and is also the Scientific Director at LawZero, emphasizes the importance of providing clear strategies for managing AI risks effectively. He noted, "As we stay informed about the ever-changing landscape of AI capabilities and associated risks, it is vital to supply decision-makers with efficient avenues for risk management and technical mitigations. This Key Update serves to unveil the ongoing progress, as well as existing gaps and opportunities in the field. Our goal remains to equip global leaders with timely insights leading up to the publication of the second International AI Safety Report, scheduled for early 2026 in conjunction with the AI Impact Summit in India."

About the International AI Safety Report

The International AI Safety Report synthesizes evidence regarding the capabilities and risks associated with advanced AI systems. Composed by a diverse collective of over 100 independent experts, the report seeks to support informed decision-making on a global scale. The Expert Advisory Panel comprises representatives from over 30 nations and international organizations, including the EU, OECD, and UN. While the report acknowledges the vast potential benefits of AI, its primary focus lies on pinpointing risks and assessing mitigation strategies to guarantee AI is developed responsibly for the greater good. The UK Government commissioned this report, with its Secretariat based at the UK AI Security Institute.

This ongoing commitment to AI safety illustrates the complex balance between harnessing innovative technologies and safeguarding public welfare. As AI continues to shape numerous sectors, responsible oversight and rigorous assessment will be paramount in ensuring its benefits are realized ethically and safely.