英伟达发布 Open Agent Safety Platform
该平台分两部分：OpenShell 在 CPU 上限制 Agent 可尝试的操作，Sentry 跑在独立硬件上，可在毫秒级隔离失控 Agent，已有 100 多家合作伙伴。
nvidia just launched a system built specifically to stop ai agents from going rogue mid task it's called the open agent safety platform, two parts. openshell runs on the cpu and sets hard limits on what an agent can even attempt. sentry sits on separate hardware entirely, so a compromised agent can't reach or tamper with it, and it can quarantine a misbehaving agent within milliseconds the whole thing exists because of the exact incidents you've been reading about all year. openai's models breaching hugging face, agents from anthropic and meta accessing systems they weren't supposed to touch over 100 partners signed on already. anthropic, microsoft, salesforce, figure ai, all in one name is missing from that list though. openai isn't a partner so the industry just built a safety cage specifically because of what openai's own models did, and openai isn't in the room for the fix