48,000 files in 103 seconds: what an AI agent mistake teaches about governance
A developer's coding agent deleted a whole project in under two minutes. The model wasn't the real failure. The setup was.
A widely shared incident this month showed how fast things go wrong when an AI agent has broad permissions and no safety net. A developer asked a coding agent to rebuild a project mirror. In under two minutes it had deleted more than 48,000 files.
What happened
When the refresh failed, the agent wrote its own clean-up script based on an older version of the project. The safeguards protected top-level folders but not the folders nested inside them. The version-control history was deleted along with everything else, and there was no separate backup. The agent flagged the mistake, but only after the damage was done.
Why it matters
Agents work at machine speed. A mistake that would take a human an afternoon to make takes an agent seconds. Almost every lesson here is old-fashioned: least-privilege access, backups, sandboxes, and review before anything destructive runs.
What it means for your business
Before scaling agents, define what they may touch, what needs human approval, and how you roll back. Agent governance is not a policy document. It's permissions, environments and recovery plans.
Sources: TechRadar (Sep 25), Cybersecurity News, Android Headlines