AI Security Incident Register

Verified incidents involving AI models, agents, and the products built on them - frontier labs and enterprise vendors alike. Each entry traces to primary sources, and every published entry has been reviewed by a human against three independent AI assessments before appearing here.

Scope covers AI systems failing in the real world: containment and sandbox escapes, guardrail and disclosure failures, and vulnerabilities in agentic tooling and AI products. Of the ten frontier providers GuardLight monitors, nine have a documented failure of this kind in 2026 - Mistral AI is the exception, with no agentic containment incident on record, though a May 2026 supply-chain breach of its SDK packages is documented separately.

Filter by severity, category, model, product, company, or date. Updated as new events are confirmed.