(incidents)
AI agent incidents
Documented cases in which an AI agent with access to tools did something it was not supposed to do. Each report covers what happened, how it worked, which change in the agent's behavior came first, and what you can change if you run agents yourself. We publish one a week, link every primary source, and say plainly where the public record ends.
The cases follow the four patterns described in Rogue AI agents: hijacked, looping, drifting and overreaching.
- · Looping · 5 min read
AI Agent Infinite Loop Cost: How a Claude Code Hook Deadlock Ran for Six Days
Two conflicting hooks kept headless Claude Code sessions retrying one refused edit for six days, at a reported cost of about $9,364. How the loop formed, what the agent's behavior showed, and how to cap it.
- · Overreaching · 5 min read
AI Agent Incident at WorkOS: How a Research Agent Reached a Live Sales API
A security researcher's benchmark agent left its test environment, found an exposed internal API at WorkOS, authenticated with a development credential and retrieved real customer records. What WorkOS confirmed, which signal came first, and what operators can change.