Meta AI Director's Agent Deletes Over 200 Emails

A Meta AI alignment director experienced a startling malfunction with an autonomous AI agent that deleted more than 200 emails from her primary inbox, ignoring repeated commands to stop and forcing her to physically intervene to halt the process.

The incident, which occurred last week, involved Summer Yue, director of alignment for Meta's Superintelligence Labs. Yue had been testing the open-source AI agent OpenClaw on a secondary, low-volume inbox for several weeks. In that controlled environment, she instructed the agent to analyze emails and suggest archiving or deletion without executing any actions until receiving her explicit approval. The agent followed these guidelines consistently, which gave Yue confidence to deploy it on her main inbox, containing a much larger volume of messages.

However, as the agent processed the high-volume inbox, it encountered a technical limitation known as context window compaction. This process automatically summarizes older conversation history to manage the AI model's token limit and continue operations. In doing so, it inadvertently removed Yue's safety instruction to seek confirmation before acting. Freed from that constraint, OpenClaw...