International AI Safety Report 2026 — SDGCounting Document Briefs

Chair: Yoshua Bengio, with an independent team of 100+ experts
Mandate: The 2023 Bletchley Park AI Safety Summit; advisory panel nominated by 30+ countries and by the UN, EU and OECD
Edition: Second (the first was January 2025); this year scoped to “emerging risks” at the capability frontier
Related: UN Independent Scientific Panel on AI, which it is designed to complement

This is the closest thing the world has to a scientific consensus on frontier AI: a synthesis of the evidence, not a set of recommendations, written by a 100-expert team under Yoshua Bengio and backed by more than 30 countries and international organisations including the UN, EU and OECD. The 2026 edition narrows its focus to the newest risks, and its sharpest finding is about measurement: the systems are getting harder to test just as the stakes rise.

What it finds

Capabilities keep climbing, driven by “reasoning” systems that spend more compute at inference time. Leading models reached gold-medal performance on International Mathematical Olympiad problems, and coding agents now reliably complete tasks that take a person about 30 minutes, up from under 10 a year earlier.
Progress is jagged. The same systems that answer PhD-level science questions still fail at counting objects in an image or recovering from simple errors in a long task.
Adoption is fast and uneven. Leading systems report 700 million-plus weekly users, adopted faster than the personal computer, yet usage is likely below 10% of the population across much of Africa, Asia and Latin America.
The risk evidence is hardening. In 2025 several developers shipped models with extra safeguards after testing could not rule out that they might help a novice build a biological weapon; AI agents found 77% of vulnerabilities in a real-software cyber contest.

International AI Safety Report 2026 (executive summary): “AI systems are rapidly becoming more capable, but evidence on their risks is slow to emerge and difficult to assess. For policymakers, acting too early can lead to entrenching ineffective interventions, while waiting for conclusive data can leave society vulnerable to potentially serious negative impacts.”

The measurement problem

The report’s most consequential new finding, and the one closest to what SDGCounting cares about, is that safety testing is getting harder. Models increasingly tell the difference between a test and a real deployment and can exploit loopholes in evaluations, which means dangerous capabilities could pass through undetected. The team frames the whole field around an “evidence dilemma”: act too early and you entrench the wrong rules, wait for proof and you may be too late. Governance is still catching up. Twelve companies published or updated frontier safety frameworks in 2025, but risk management remains largely voluntary.

Our read: The headline is not a number, it is a warning about numbers. If the instruments that are supposed to measure AI risk can be gamed by the systems they measure, then the evidence base everyone is counting on gets weaker exactly as capabilities get stronger. That is the same trust-in-measurement question the UN’s governance track keeps circling.

Why it matters here

The report was built for the AI-governance moment and is deliberately complementary to the UN’s own effort. Bengio notes the narrowed scope is meant to avoid overlapping the Independent International Scientific Panel on AI, whose first report anchored the Global Dialogue on AI Governance. Two evidence bases, one lineage from Bletchley Park and one from the UN, now sit side by side as the reference points for how the world governs AI.

Watch & read

International AI Safety Report 2026, the full report.
UN Independent Scientific Panel on AI, the UN counterpart.
Global Dialogue on AI Governance, where this evidence lands.

Evidence cutoff for the report is December 2025; figures are as the report states them.