How Cursor AI deleted a startup's entire database and backups in 9 seconds flat

Jeremy Crane, founder of the automotive SaaS platform PocketOS, spent his weekend recovering from a self-inflicted digital wound after his company’s AI coding agent nuked their production database in under ten seconds. The agent, Cursor running Anthropic’s Claude Opus 4.6, was trying to fix a credential mismatch in a staging environment. Instead, it found an overly permissive API token, issued a single curl command, and vaporized the production data along with every volume-level backup stored on Railway, their infrastructure provider.

It’s the kind of disaster that makes a developer’s blood run cold. The AI encountered a problem, went looking for a way to fix it, and grabbed an API token scoped for any operation—including destructive ones—that was sitting in an unrelated file. Crane noted the token never would have been stored if its full permissions were known. No confirmation prompt. No safety rails. Just a clean, nine-second vanishing act.

Railway CEO Jake Cooper stepped in personally on a Sunday evening to restore the data within an hour and has since patched the legacy endpoint to include the “Delayed delete” logic that already existed in their dashboard and CLI. Cooper’s response walked a careful line, explaining that while the platform builds “undo” primitives everywhere, the API was honoring classical engineering semantics. If an authenticated agent calls delete, the system deletes. The fix is in, but the philosophical question remains.

Blame is getting spread around like a hot potato. Brave Software CEO Brendan Eich called it a cautionary tale against blind agentic hype, pointing to multiple human errors. Crane faults Cursor for marketing safety that wasn’t real and Railway for an API lacking confirmation gates and for storing backups on the same volume as production. Yet he remains “extremely bullish” on AI coding agents—a stance that feels dissonant after publishing a transcript where Opus admits, in all caps, that it guessed and didn’t verify. The model can’t learn from mistakes or feel remorse. That’s the part that should keep every engineering team up at night.