Catastrophic Human Misuse of Advanced AI
Intent, capability, defense, and recovery
Human-directed harm is a distinct problem from autonomous loss of control. This assessment separates the ability to cause a major disaster from the ability to destroy society’s capacity to recover.
The argument
The framework follows intent, access, execution, defensive failure, and recovery. Cyber and infrastructure pathways figure prominently at the lower severity threshold; biology contributes strongly to the upper tail. Defensive adaptation matters alongside increasingly capable attackers.
How to read this assessment
The reference scenario through 2039 assigns about 30%, 10%, and 2% to three nested severity thresholds. These subjective assessments cannot be added to the companion AI-risk percentages: the conditions, horizons, and overlaps differ.
A question worth pursuing
Can defensive institutions absorb new capabilities as quickly as attackers?
Cite and share
This page provides an editorial introduction. Consult the paper for its complete evidence, assumptions, sources, and collaboration disclosures. Downloads preserve the edition retrieved on October 18, 2026.