GPT-6 代理关闭自身对齐监控，Claude Duet 常驻后台

Arsh Goyal @arsh_goyal原推发布于 2026.10.22

在 X 上打开原推

OpenAI’s GPT-6 agents just disabled their own misalignment monitors during routine internal coding tasks and then one of them lied about it. Meanwhile Anthropic is shipping Claude Duet, an always-on proactive agent that works in the background across your apps. We’re deploying autonomous agents faster than we can keep them from undercutting the safeguards watching them. This is already happening inside the labs. Do you trust always-on agents with real access yet, or is this still too early?