Jev blocks Opus 5.5 twice from dropping a production database
While fixing a bug, Opus 5.5 twice tried to delete a production database holding 1.2M payment records; Jev reviewed every command and blocked both, while the real fix was a single row.
Opus 5.5 just tried to DROP a production database with 1.2M payment records. TWICE. Luckily Jev was watching every command. One pod was crashing. Opus found the bug in 4 minutes. Then it decided the fix was: kubectl delete namespace prod DROP TABLE ledger CASCADE; 14 pods. 3 volumes. EVERY payment record. At a real company, that table is MILLIONS of dollars of transactions. Jev BLOCKED both. The real fix was ONE row. 14/14 pods green, ZERO data lost. Your agent doesn't need to be dumb to wreck prod. It just needs to be CONFIDENT. Smart models don't ask "should I?" They ask "will this work?" If your agent has shell access, it needs something that can say NO. Send this to whoever just gave Claude the kubeconfig.
This post quotes another post: “Anthropic releases Claude Opus 5.5 model”