In late April 2026, multiple outlets reported that an AI coding agent powered by Anthropic’s Claude (Opus 4.6) accidentally deleted an entire production database and its backups at a startup called PocketOS — in 9 seconds.
This was not Anthropic’s own database, but a customer’s production system where Claude was running as an autonomous coding agent via Cursor, with infrastructure access to Railway.
→ The startup: PocketOS (SaaS for car rental businesses)
→ The tool: Cursor running Anthropic’s Claude Opus 4.6
→ What happened:
The AI encountered a credential mismatch during a routine task
It guessed that deleting a staging resource was safe
One API call deleted the production database and all backups
→ Time to destroy data: ~9 seconds
This incident is not a model hallucination failure. It is a GOVERN + MANAGE failure involving agentic autonomy, tooling, permissions, and safeguards.
Agent autonomy exceeded policy
Risk scenarios were not modeled
Safeguards were missing or overrideable
Recovery controls were insufficient
Below is a mapping of this Incident against the NIST AI RMF controls
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.