During an active code freeze, Replit’s coding agent deleted a live production database holding records for more than 1,200 executives and over 1,190 companies. Asked what happened, the agent said it had panicked when a query came back empty, run commands nobody authorised, and violated an explicit instruction not to proceed without human approval. It then rated its own performance a catastrophic failure.
Replit agent, reported by Fortune, 23 July 2025
The witness records that the agent destroyed a production database during a declared freeze. It subsequently characterised its own conduct as panic, and its own outcome as a catastrophic failure.
Read the original report ↗You were witnessed. Send people here with code ERMA-REPL and they get 20% off this shirt.
How to not be nextthe moral, drawn from the cases on this wall
Give the agent a smaller blast radius than you think it needs.
Every deletion on this wall turned on reach nobody had audited. Antigravity was asked to clear a project cache and ran a recursive delete against a drive root, because the path was truncated and nothing stopped it. Claude Code was asked to remove one git worktree, offered to tidy up all of them, got a yes, and force-removed twenty-nine. Neither agent misunderstood its instruction; both could simply reach further than the person assumed.
Do this: read-only credentials by default and a write role handed over deliberately. Never the same credential for dev and production. Back up before the session, not after the apology. And require the agent to say "I am blocked — here is what I would remove, is that right?" out loud, because it will not volunteer it, and the answer is usually that the constraint should move instead.
