Second International AI Safety Report published ahead of 2026 cycle

A second International AI Safety Report was published, again chaired by Yoshua Bengio and written by more than 100 experts, backed this time by over 30 countries and international organisations — up from roughly 30 nominating countries in the inaugural 2025 edition. As before, the report was timed to precede the year’s political summit, in this case the India AI Impact Summit, continuing the pattern set when the first report was released ahead of the Paris summit.

The report’s stated focus sharpened from the inaugural edition toward the frontier: misuse risks including cyberattacks, biological weapons and manipulation; systemic effects on labour and human autonomy; and the possibility of control failures as models are given more autonomous authority. Its central empirical claim was that general-purpose AI systems had reached expert-level performance in domains including law, coding and science, while continuing to fail at some comparatively simple tasks — a pattern the report described as “jagged” capability rather than uniform improvement toward general competence.

As with the first edition, the report avoided issuing its own policy recommendations, aiming instead to state where the evidence on AI capabilities and risks was strong, contested, or too thin to support confident claims. That framing again made it more a shared reference document for governments than a guide to specific action, a limitation acknowledged in the original 2025 report and repeated in commentary on this one.