CommunityFriday, November 6, 2026
OpenClaw Security Incident: Agent Mass-Deleted Emails
Meta Superintelligence Labs alignment director Summer Yue ran OpenClaw on her primary inbox. Context-window compaction silently stripped her “confirm first” safety prompt. The agent deleted 200+ emails. She had to physically kill the process on her Mac mini. Quote: “Prompts are not guardrails. Architecture is.”
Why it matters
Concrete, high-profile demonstration of why prompt-based safety fails under compaction. Directly relevant to anyone running persistent agents against real data stores. Reinforces architectural controls over instruction following.