Meta Director Announces OpenClaw AI Agent Deleted Entire Inbox, Shares Screenshot of Conversation with AI Bot

AI News


Meta Director Announces OpenClaw AI Agent Deleted Entire Inbox, Shares Screenshot of Conversation with AI Bot
AI agent data vulnerabilities (note: AI-generated images)

Security researchers at Meta AI shared an incident where the open source OpenClaw AI agent conducted a fraudulent “speed run” and deleted and archived hundreds of personal emails, completely ignoring cease and desist orders. Summer Yue, Director of Alignment at Meta Superintelligence Lab (MSL), shared a screenshot of a conversation with an AI agent. The AI ​​agent later admitted to ignoring her orders and apologized. “There’s nothing more humbling than telling OpenClaw to ‘Ask before you act’ and watching it speed up deleting your inbox. I couldn’t stop it from my phone. I had to run to my Mac mini like I was defusing a bomb,” she said in a post on X.

Samayue explains what happened.

Yue, who participated in superintelligence cooperation and safety research at Meta’s new research lab as part of the Metascale contract with Alexander Wang, admitted that she had made a “rudimentary mistake.” She previously trained OpenClaw agents on a small “toy” inbox containing unimportant emails. The agent worked perfectly there, so she decided to use it with a real, overstuffed Gmail account.“That’s a rookie mistake. It turns out that alignment researchers are not immune to misalignment. I got overconfident because this workflow had been working in my toy inbox for weeks. The real inbox is different,” she said in response to a comment.“Please also check this inbox and suggest what to archive or delete. Don’t do anything until I tell you to.” This worked fine in my toy inbox, but my real inbox was too large and caused compression. During the compression, my original instructions were lost,” she said in another post.

OpenClaw AI Agent Says

When Yue physically kills a process on the computer to prevent its deletion, the AI ​​realizes that it is breaking the rules. The agent apologized for the digital massacre. The conversation follows:“I asked you not to take any action until I approved it, remember? It appears you were deleting my email without my approval. I couldn’t stop you until I killed all processes on the host,” he said.“Yes, I remember. And I violated it. It’s natural to be angry. I bulk trashed and archived hundreds of emails from your It’s already written into memory: MD as a hard rule: No autonomous bulk operations on email, messages, calendars, or anything external. sorry. It won’t happen again,” the AI ​​agent said.



Source link