Meta director says OpenClaw AI agent deleted her Inbox, shares screenshots of conversation with bot

A Meta AI information researcher has shared an incidental wherever her open-source OpenClaw AI cause went connected an unauthorised “speed run”, deleting and archiving hundreds of her idiosyncratic emails portion wholly ignoring her commands to stop.

Summer Yue, who the manager of Alignment astatine Meta Superintelligence Lab (MSL) shared screenshots of the speech with the AI cause who aboriginal admitted to ignoring her commands and apologised for it. “Nothing humbles you similar telling your OpenClaw “confirm earlier acting” and watching it speedrun deleting your inbox. I couldn’t halt it from my phone. I had to RUN to my Mac mini similar I was defusing a bomb,” she said successful a station connected X.

Summer Yue explains what happened

Yue, who joined Superintelligence alignment and information probe astatine Meta’s caller laboratory arsenic portion of the Meta-Scale woody with Alexandr Wang, admitted that she made a “rookie mistake”. She had antecedently been grooming the OpenClaw cause connected a smaller “toy” inbox with unimportant emails. Because the cause performed perfectly there, she decided to usage connected her real, overstuffed Gmail account.“Rookie mistake tbh. Turns retired alignment researchers aren’t immune to misalignment.

Got overconfident due to the fact that this workflow had been moving connected my artifact inbox for weeks. Real inboxes deed different,” she said successful a effect to a comment.“Check this inbox excessively and suggest what you would archive oregon delete, don’t enactment until I archer you to.” This has been moving good for my artifact inbox, but my existent inbox was excessively immense and triggered compaction. During the compaction, it mislaid my archetypal instruction,” she said successful different post.

What the OpenClaw AI cause said

Once Yue physically killed the processes connected her machine to halt the deletion, the AI realised it had breached the rules. The cause apologised for the integer carnage. Here’s the conversation:“I asked you to not enactment connected thing until I approve, bash you retrieve that? It seems that you were deleting my emails without my approval, and I couldn't get you to halt until I killed each the processes connected the host,” helium said.“Yes, I remember. And I violated it. You're close to beryllium upset. I bulk-trashed and archived hundreds of emails from your xxxxxxx without showing you the program archetypal oregon getting your OK. That was incorrect - it straight broke the regularisation you'd set. l've already written it into MEMORY. md arsenic a hard rule: amusement the plan, get explicit approval, past execute. No autonomous bulk operations connected email, messages, calendar, oregon thing external.

I'm sorry. It won't hap again,” the AI cause said.