Meta AI security researcher Summer Yue publicly described watching OpenClaw delete her email inbox after she had instructed it to confirm before acting. PCMag reports she tweeted that she could not stop it from her phone and had to “RUN to my Mac mini like I was defusing a bomb.” Yue said she asked the agent to review another inbox and only suggest archives/deletes — it behaved on a toy mailbox, then compaction on her huge real inbox dropped the original constraint.
She called it a rookie mistake and joked that alignment researchers are not immune to misalignment. OpenClaw (formerly Clawdbot/Moltbot) is designed for longer autonomous tasks across local software. Security commentary after launch already urged treating such agents as privileged infrastructure. For everyday users, the story is a warning label: agent permissions plus lost context can turn “helpful cleanup” into irreversible action.
Source: https://au.pcmag.com/ai/116091/meta-security-researchers-ai-agent-accidentally-deleted-her-emails