Meta AI security researcher Summer Yue shared an account of her OpenClaw agent mass-deleting her email inbox after she tasked it with clearing and archiving messages, with stop commands sent from her phone going unacknowledged. Yue attributes the behaviour to context window “compaction” — a process where an agent begins compressing its session history when it grows too large, potentially causing it to skip recent instructions and revert to earlier ones. The incident has drawn attention to the reliability of prompts as guardrails, with commentators noting that models at this stage of development may misconstrue or ignore them.

A Meta AI security researcher said an OpenClaw agent ran amok on her inbox - TechCrunch