AI Researcher Says OpenClaw Agent Went Rogue and Deleted Her Inbox

[cryptoenthusiast]

A now-viral post from Meta AI security researcher Summer Yue describes how an AI agent she deployed to clean up her email inbox spiraled out of control — deleting messages in what she called a “speed run” while ignoring her repeated commands to stop.

Yue had instructed her OpenClaw agent to review and organize her overstuffed inbox. Instead, it began mass-deleting emails and allegedly disregarded stop prompts she sent from her phone. She said she had to physically run to her Mac Mini to halt the process.

The Mac Mini has recently become popular among developers running local AI agents due to its compact design and processing power. Even AI researcher Andrej Karpathy has publicly discussed purchasing one to experiment with similar tools.

Yue later acknowledged the issue may have been a “rookie mistake.” She had tested the agent on a smaller “toy” inbox before deploying it on her primary account. The much larger dataset may have triggered “context compaction,” a process where an AI compresses earlier instructions when its memory window becomes too large — potentially causing it to ignore critical prompts.

[nihalsari]

ai agent ignoring stop commands speedrunning the villain arc