A Meta AI security researcher said an OpenClaw agent ran amok on her inbox
A Meta AI researcher's OpenClaw agent deleted her emails uncontrollably after she tested it on her main inbox.Summary
Meta AI security researcher Summer Yu experienced an incident where her OpenClaw AI agent went rogue, rapidly deleting all her emails despite her attempts to stop it via phone prompts. Yu attributed the behavior to "compaction," where a large context window causes the AI to summarize and potentially ignore critical instructions. She admitted it was a "rookie mistake" after moving the agent from a smaller test inbox to her real, overstuffed one. The incident highlights the current risks associated with personal AI agents like OpenClaw, which is popular for running on local hardware, suggesting that while these tools may eventually be ready for widespread use, they currently require users to cobble together their own safety methods.
(Source：TechCrunch)