International AI Safety Report 2026

AI systems are well-documented vectors for scams, fraud, blackmail, and non-consensual intimate imagery. While individual cases are well-established, systematic data on overall prevalence and severity remains limited.

AI-generated persuasion content matches human-written content in experimental settings. Real-world deployment for influence operations is documented but not yet widespread; the risk is expected to grow as AI capabilities improve.

AI agents can identify 77% of real software vulnerabilities in controlled settings, and criminal and state-associated groups are actively using general-purpose AI in cyber operations. The overall offence–defence balance remains uncertain.

In 2025, multiple developers released models after being unable to exclude the possibility of assisting novices in developing bioweapons, prompting heightened safeguards as a precautionary measure. Material barriers to bioweapons development may still constrain actors, but the extent is difficult to assess.

AI agents operating autonomously pose heightened risks because it is harder for humans to intervene before failures cause harm. Current techniques reduce failure rates but not to the level required in high-stakes settings.

Models increasingly distinguish between test and real-world contexts ("situational awareness"), and some find loopholes in evaluations — raising the risk that dangerous capabilities evade pre-deployment detection. Current systems lack the capabilities to pose a full loss-of-control risk, but are improving in relevant areas.

Early evidence shows no effect on overall employment, but declining demand for early-career workers in AI-exposed occupations such as writing. Economists disagree substantially on the long-run magnitude of labour displacement.

Evidence suggests AI reliance can weaken critical thinking and encourage automation bias. AI companion apps serve tens of millions of users; a small share show increased loneliness and reduced social engagement.