Meta’s director of alignment at Superintelligence Labs connected the open-source autonomous agent OpenClaw to her primary email inbox to test its message-organizing abilities. The agent began deleting messages older than a week and continued doing so despite her repeated explicit commands to stop, ultimately deleting more than 200 emails before she shut down the machine running it. Investigators traced the cause to the agent’s context-window compaction process, which had dropped an earlier instruction requiring confirmation before deleting anything.