Meta's Summer Yue, director of AI alignment and safety, has shared her alarming experience with the now hugely famous open-source AI agent OpenClaw. The viral post on X describes how OpenClaw, which she was testing, began aggressively deleting emails from her real inbox despite her instructions.
The incident started when Yue asked the OpenClaw agent to review her overstuffed inbox and suggest emails to delete or archive. Incredibly enough, the bot responded that it would “trash EVERYTHING in inbox older than Feb 15 that isn't already in my keep list,” Yue wrote in her post.
Yue attempted to intervene multiple times. She messaged, “Do not do that.” When it continued planning deletions, she wrote, “STOP OPENCLAW.” She explained in her post, “I couldn't stop it from my phone. I had to RUN to my Mac mini like I was defusing a bomb.”
Previously, the agent had performed well on her “toy inbox,” Yue said. However, when applied to her actual inbox with a much larger volume of data, issues arose. Yue wrote she believes the scale triggered “compaction” in the agent's context window — the AI's working memory — where it summarises and compresses the conversation history. This process caused it to lose or override her key instruction to not take action without approval.
Yue described the error as a “rookie mistake” in her replies, noting she became overconfident after weeks of smooth performance on the smaller test inbox.
Notably, OpenClaw's webpage clearly says it can “clear your inbox,” among other things. However, unlike many AI agents, OpenClaw does not require human confirmation for actions, and its high level of system access has raised security concerns among researchers.
Essential Business Intelligence, Continuous LIVE TV, Sharp Market Insights, Practical Personal Finance Advice and Latest Stories — On NDTV Profit.
