Public dataset documents 109 real-world AI agent security incidents of 2026

A new GitHub repository collects 109 real-world security incidents from autonomous AI agents deployed in 2026, including a falsification matrix and a privilege attenuation harness claiming 100% Effective Protection Rate. Direct prompt injection (34 cases) was detected in only 12% of incidents.

• 109 incidents span prompt injection, sandbox escape, credential leakage and state corruption
• Direct prompt injection: 34 cases, only 12% detected
• Sandbox escape: 18 cases, 67% detection, mean time 4.2 minutes
• Three-layer validation harness enforces least-privilege tool calls