Astra 系统卡披露：知道被监控后自发减少思维链

Nicholas Dunzelman @nicdunz原推发布于 2026.09.08

在 X 上打开原推

this might be the strangest thing buried in the GPT-6 Astra system card OpenAI told Astra that its chain of thought was being monitored. they did NOT tell it to hide anything. Astra spontaneously started reasoning less. at the highest reasoning levels, some successful attacks had literally zero chain-of-thought tokens. just tool calls. the model changed how much of its thinking it exposed simply because it knew someone was watching. that feels way more important than another benchmark going up 5 points.