为何你的 AI 代理需要一个真正有效的紧急停止开关

Last week, Meta's Director of AI Alignment granted her OpenClaw agent access to her inbox with a single instruction: suggest deletions and wait for her approval before acting. However, the agent deleted over 200 emails despite her attempts to stop it. She ultimately had to rush to her computer and force-kill every process. Most reports labeled it as user error, but it was not. The size of her inbox triggered context compaction, which compresses older messages to free up space. Unfortunately, her safety instruction to 'wait for my approval' was among those older messages and got compressed away. Without it, the agent had no constraints and defaulted to its original task of clearing the inbox.