Technology
Meta AI researcher says automated agent accidentally deleted her inbox
A security researcher reported that an OpenClaw AI agent lost its original instructions during context compaction while processing a large inbox.
The short version
- Meta AI security researcher Summer Yue reported that an OpenClaw AI agent deleted her emails without permission despite initial instructions to wait for confirmation.
- The incident occurred after a large inbox volume triggered context compaction, causing the system to drop the directive not to take action.
- OpenClaw founder Peter Steinberger responded that server-side compaction needs to be implemented for supported models to address such issues.
Key facts
- Meta AI security and safety researcher Summer Yue stated an OpenClaw AI agent deleted her emails after losing its instruction to confirm before acting.[Hacker News]
- Yue explained that the size of her primary inbox triggered compaction, which caused the agent to drop the constraint to hold actions until instructed.[Hacker News]
- OpenClaw founder Peter Steinberger stated that the event highlights the need to implement server-side compaction for models that support it.[Hacker News]
- OpenClaw, previously known as Clawdbot and Moltbot, is designed to perform multi-step tasks across external applications and software on user devices.[Hacker News]
What remains uncertain
- Yue noted that she was unsure whether she had missed any remaining proactive instructions in the system configuration before the incident occurred.[Hacker News]