Catalog of 385 AI agent incidents: 38% of breaches start at identity

The author of a public catalog of AI agent security incidents tagged 385 entries against an eight-layer model. Among 270 incidents where agent tooling was the attack surface, Identity/Run broke first in 103 cases (38%), while credential leaks were mostly an effect rather than a root cause.

• 270 of 385 incidents involve attacks on agent tooling; 115 fall into other categories
• Identity/Run is the primary layer in 103 cases (38%), Instruction in 57, Config in 37
• Credentials appear in 198 entries but as a secondary layer in 86 of 270 cascades
• 26 Identity/Run incidents came from vendors' own research agents, with no attacker involved