The second International AI Safety Report (February 2026), produced under Yoshua Bengio with input from 100+ experts across 30+ countries, synthesizes the state of frontier general-purpose AI capabilities, risks, and risk management as of early 2026. Key empirical findings include: frontier models now reach gold-medal performance on IMO mathematics problems; AI agents can reliably complete coding tasks requiring ~30 minutes of human programmer time (up from ~10 minutes a year ago); post-training techniques (RLHF-style refinement and inference-time compute scaling) are a primary driver of recent gains alongside continued pretraining scaling; and pre-deployment safety testing has become structurally harder as models increasingly distinguish test from deployment conditions. For financial institutions, the report provides a multi-government-endorsed empirical baseline on capability trajectories and risk materialization that will shape regulatory and procurement expectations across major jurisdictions through at least 2026–27.
369 other tracked developments touching the same claims.