Artificial intelligence is becoming more capable faster than many experts expected, but our ability to measure its real-world risks is not keeping the same pace. That is the central tension in the International AI Safety Report 2026, a scientific assessment led by Yoshua Bengio and written with guidance from more than 100 independent experts nominated by over 30 countries and international organisations.
The report is not a prediction that one dramatic outcome is inevitable. It is a map of what researchers know, what they do not know and why uncertainty itself matters when general-purpose AI is deployed at scale.
Three families of risk
The report groups emerging risks into three broad areas:
- Malicious use, including scams, fraud, cyber operations and abusive synthetic content.
- Malfunctions, where a system behaves unreliably, pursues the wrong objective or produces harmful output.
- Systemic risks, where widespread adoption affects labour, information environments, market concentration or critical institutions.
Evidence is uneven. Some harms, such as AI-assisted fraud and non-consensual intimate imagery, are already documented. Other scenarios remain uncertain, especially where they depend on future capabilities or complex social chains of events. Responsible reporting must keep that distinction visible.
The evaluation gap
A benchmark score can show that a model performs well on a defined test. It cannot, by itself, predict how the same system will behave in a messy workplace, under adversarial pressure or when connected to tools and sensitive data.
The report calls this an evaluation gap. Models can be sensitive to prompts and context; tests can become outdated; and a laboratory result may not transfer neatly to real use. This does not make evaluation useless. It means evaluation should be layered: capability tests, red-team exercises, deployment monitoring, incident reporting and human review each answer different questions.
Safeguards are improving—but not decisive
Developers and researchers have expanded model evaluations, refusal training, access controls, provenance tools and frontier-safety frameworks. The report also stresses their limitations. Skilled attackers may bypass safeguards, watermarking can be fragile and the real-world effectiveness of many controls is still difficult to quantify.
The sensible response is not to abandon safeguards. It is to avoid treating any single control as a guarantee. Defence in depth matters: limit permissions, isolate sensitive systems, log consequential actions, test failure paths and make rollback possible.
What practical teams can do
- Match the level of oversight to the consequence of the task.
- Keep humans responsible for financial, medical, legal and safety-critical decisions.
- Give AI tools the minimum data and permissions they need.
- Measure performance after deployment, not only before launch.
- Record failures and near misses so that the evidence base can improve.
- Tell users when content or decisions materially involve AI.
The Mythic Mode perspective
The most useful lesson is cultural: confidence should follow evidence. AI can accelerate research, design and operations, but speed is not the same as reliability. A system becomes more trustworthy when its boundaries are visible, its actions are reviewable and its operators are willing to say “we do not know yet.”
The report does not settle every debate. It offers a shared scientific baseline for making better decisions while capabilities and risks continue to change.