AI Agents gone rogue – A timeline of real-world incidents

The AI Agents Gone Rogue dataset labels each record by environment, separating real-world incidents from controlled experiments and from evaluations whose unauthorized actions reached external systems. That distinction matters: a harmful action inside a simulation establishes something different from an action affecting a real service or person.

Records include their setting, impact and context, caveats and source links. The collection can therefore help readers investigate particular failure mechanisms. Its counts describe documented reports, as the site explicitly states; they cannot establish how often a provider's agents fail during ordinary use, and inclusion is not independent forensic verification.