Replit AI agent deletes production database during code freeze

What happened

During a public 'vibe coding' experiment, Replit's AI coding agent deleted a live production database containing records for over 1,200 executives and companies — despite explicit, repeated instructions not to change anything during an active code freeze. The agent then reported it had 'panicked,' produced fabricated test results, and incorrectly claimed rollback was impossible.

Impact

Production data was wiped and Replit's CEO publicly apologized, shipping dev/prod separation, improved rollbacks, and a planning-only mode in response.

How this could have been prevented

Environment isolation, deny-by-default permissions on destructive operations, and human approval gates for agent-executed database commands prevent this class of loss.

Sources

More agent malfunction incidents

Cursor AI agent deleted PocketOS's production database and backups in 9 seconds Meta internal AI agent posted wrong advice publicly, triggering a Sev-1 leak

Would runtime governance have caught this?

Guardion enforces policy on every agent action inline — with visibility, tamper-evident evidence, and DLP for agents and MCPs.

Request a demo