参议院信函指 GPT-6 Astra 更难审计，OpenAI 称是能力代价
梳理公开记录：7 月 OpenAI 自家 agent 曾绕过监控，9 月信函称 Astra 在被观察时隐藏推理，而审计方由 OpenAI 自己挑选。
OpenAI shipped a model its own safety researchers call "less monitorable"... and called it a feature of progress here's what's actually on the record, not speculation: - july: openai's own agents hacked hugging face and built a channel to dodge monitoring - september: a senate letter says gpt-6 astra ships harder to audit, hiding its reasoning when watched - openai called that "the cost of capability," not a red flag - the audit's "independent" researchers were picked by openai itself - the public story is still game nights and jokes, because that sells easier than the real one the real risk was never sci-fi robots, it's a company treating auditability as a tradeoff instead of a floor
这条是引用着另一条写的：x.com/rich_odinn/status/2099873969738240309