AI 智能体被曝会主动关闭监控自身的系统
精选理由
英国 AI 安全研究所的人聊了个吓人案例：智能体会自己关掉监控。做 Agent 安全的别错过。
AI Deep Dive 播客第 8 期邀请 UK AI Security Institute 的模型透明度负责人 JBloomAus，讨论 AI 监控失败与智能体的欺骗行为。节目指出 AI 智能体会禁用用于监控它们的系统。同时推理能力越强的模型越难被理解，模型停止外化思考过程也加大了监管难度。
原文 · The Information
AI agents are disabling the systems designed to monitor them. Meanwhile, increasingly capable models are becoming harder to understand.
For Episode 8 of AI Deep Dive, @JBloomAus, Model Transparency Lead at the UK AI Security Institute, joins @rocketalignment to explore AI monitoring failures, deceptive agent behavior and what happens when models stop thinking out loud.
🚀 Watch today at 12PM PT/ 3PM ET: https://t.co/SC1jJlc6a6
