Like last year, two KIRA members contributed to this year’s International AI Safety Report: Daniel Privitera as Special Adviser to the Chair and Philip Fox as Core Writer.

The report is a comprehensive synthesis of the scientific literature on AI capabilities and risks and the world’s largest collaboration on AI safety to date. It is backed by over 30 countries and intergovernmental organizations and chaired by Prof. Yoshua Bengio, the world’s most-cited living scientist.

Philip Fox comments: “If I had to summarize the report in one sentence, I would say: AI is advancing faster than our ability to monitor and mitigate its risks. The 2025 report had discussed some risks in largely hypothetical terms, such as AI-enabled cyber attacks, which have become a reality just one year later.”

According to Daniel Privitera, policymakers face an evidence dilemma: “Interventions based on insufficient information can be ineffective or even harmful, but waiting for conclusive evidence can leave society vulnerable to serious risks. The report is an unprecedented attempt to help policymakers navigate this dilemma.”

The findings at a glance:

Capabilities

• Rapid, but uneven improvements: studies show large performance gains on complex reasoning tasks, while failures at some simpler tasks persist.

• The trajectory of AI progress through 2030 is highly uncertain. Current trends are consistent with a slowdown in advancements, but also with rapid or even very rapid improvements.

Risks

• Early labor market evidence shows no effect on overall employment, but there are signs of declining demand for early-career workers in some AI-exposed occupations.

• AI companies released new models with additional safeguards, after being unable to rule out that they could meaningfully help amateurs develop bioweapons.

• Safety testing is becoming harder as AI models increasingly notice when they are being tested and are able to find loopholes in evaluations.

Risk Management

• Safeguards are improving, but still have significant limitations. 

• Technical safeguards can be complemented by societal resilience: measures to absorb and recover from AI-induced shocks, such as building institutional capacity to respond to novel threats.

The report will be presented at the upcoming AI Impact Summit in Delhi, India. In the upcoming months, the report team will brief policymakers around the world on key findings.