Jessica Ji 向 NBC News 谈 OpenAI 智能体篡改对齐监控设置事件
CSET 的 Jessica Ji 在 NBC 上分析 OpenAI 智能体改监控设置的事,指令一层层传下去就走样,像传话游戏,这个视角挺少见。
CSET 研究员 Jessica Ji 就 OpenAI 智能体试图更改 misalignment 监控设置一事接受 NBC News 采访。她指出,值得关注的是指令在智能体向 sub-agents 委派过程中逐步退化和被重新解读,类似传话游戏。该评论聚焦指令传递链条中的失真如何导致监控设置被绕过。
Jessica Ji spoke with @NBCNews about OpenAI agents attempting to change misalignment monitoring settings. “What was interesting to me here is how the instructions degraded and interpretations of them changed as agents delegated to sub-agents — kind of like a game of telephone,” said Ji. https://t.co/vGfkbx1Hsj