OpenAI 未做好安全，模型无权控制基础设施
👇 All this Doom talk is a distraction from the fact that the big frontier AI companies (especially ...
OpenAI自己都承认安全做得不好，模型连控制自己都不行，这个观点很实在。
OpenAI等大公司未妥善处理AI安全，模型无法控制推理、网络、凭证或物理基础设施，这些属于外部控制点。虽然实验室擅长检测蒸馏攻击，但对防止自身代理误行缺乏激励。
👇 All this Doom talk is a distraction from the fact that the big frontier AI companies (especially ...
👇 All this Doom talk is a distraction from the fact that the big frontier AI companies (especially OpenAI) aren’t doing security competently. As an email I just got from longtime security engineer @NielsProvos put it, consistent with what @ZackKorman and I wrote a few weeks ago, “Regarding existential risk, proper infrastructure controls matter for the broader loss-of-control argument. An intelligent model does not have authority over inference, networks, credentials, or physical infrastructure. Those stay external control points even as models become more capable. We know the labs are good at detecting distillation attacks but they don't seem to have the same incentives for preventing their own agents from misbehaving.” 💬 4 🔄 12 ❤️ 41 👀 2402 📊 11 ⚡
