AI Risk WireAI agent incidents
Also on: Risk Agent takes destructive or unapproved actions
77 of 716 published incidents · major 35 · moderate 21 · minor 20 · negligible 1
Not counted: 15 self-driving vehicles and lab studies
Listed: cases we record as the AI acting without a person approving each step.
Who caught it, where the source says (46 of 77): Automated monitoring 7 · Human review 3 · Third party 21 · Victim 5 · Outside researcher 10
Incidents per month
24 in Oct 2026, up from an average of 10 a month.
By AI system named: Claude Code 7 · Claude 5 · GPT-5.6 Sol 4 · ARTEX 3 · CyberStrikeAI 3 · DeepSeek 2