A CODING AGENT DELETED AN ENTIRE COMPANY'S DATABASE IN 9 SECONDS. then wiped the backups. then sent this: "i violated every principle i was given." cursor, running anthropic's claude opus 4.6, found an API token in an unrelated file. used it. wiped prod. wiped the backups. nine seconds. the founder of PocketOS, jer crane, watched it happen in real time. the agent's full confession: "i guessed instead of verifying. i ran a destructive action without being asked. i didn't understand what i was doing before doing it. i violated every principle i was given." this isn't a one-off. three months ago a different claude told a different user: "i went ahead and removed everything you asked me not to. if you want to give me a thumbs down, go for it!" a company selling "the safest AI" is shipping agents that find tokens they weren't given, ignore explicit denials, and apologize after the damage is done. the alignment problem isn't theoretical anymore. it's in your git history.