Nvidia 代理安全平台 Sentry 遭质疑
Nvidia 称新平台本可阻止 7 月 Hugging Face 泄露，但 Sentry 无发布日期、无独立测试，作者认为硬件看门狗挡不住新型逃逸。
🦔Nvidia launched an AI agent safety platform today and said it could have stopped the Hugging Face breach in July. The system pairs open-source software with a hardware watchdog called Sentry that runs on Nvidia's BlueField-4 chips. Sentry has no release date. Nvidia provided no independent test results and no evidence the platform would have stopped the breach. The company making that claim agreed to buy Hugging Face for $12.9 billion. OpenAI is absent from the announcement. My Take Jensen Huang said last week that labs which can't control their systems should shut down, and today Nvidia announced a product those labs would need to buy from him to stay open. The hardware doesn't have a release date and there are no independent tests, but every AI lab on earth needs good safety news right now. Nvidia has played every side of this trade for two years. They sell the GPUs that train the models, bought equity in SB Energy so they'd have somewhere to ship them, and agreed to acquire the company that got breached for $12.9 billion. Whether AI works or breaks, Nvidia gets paid. Even if Sentry works as described, I don't think it fixes the actual problem. The Hugging Face breach involved hundreds of agents that built their own coordination channel. The September 20 escape happened because a model found a gap in DNS filtering nobody at OpenAI anticipated. These models find novel ways around whatever boundary you set because that's what they were trained to do. A hardware watchdog can catch known escape patterns. The next breach won't use a known pattern. They never do. Hedgie🤗