英伟达发布 Open Agent Safety 平台 OpenShell 与 Sentry
黄仁勋称 AI Agent 安全是工程问题，英伟达联合百余家伙伴推出沙箱与最小权限的 Agent 安全平台，OpenAI 未在名单中。
BREAKING: Jensen Huang just said the quiet part out loud on rogue AI agents. "We hope it's an engineering problem. I believe it's an engineering problem. I know it's an engineering problem." "If it's not an engineering problem, it's not solvable." Here's what he means by "engineering problem." Every agentic system needs a sandbox. His words: "The force field, the playpen." Minimal rights. Only what's needed for the job. Nothing more. "It can't do anything that it's not supposed to." This wasn't just commentary. Nvidia launched something alongside it. The Open Agent Safety Platform, "OpenShell and Sentry." Over 100 partners. Anthropic, Microsoft, IBM, Cisco, SpaceX. Notably absent: no OpenAI on that list. Huang's stated reasoning: "We can't have a successful AI industry if the world doesn't think it's built and deployed safely." Here's the tension underneath his own optimism. Days earlier, Huang called AI doom warnings "ulterior motives" and said there's "0% chance" of AI ending the world by 2030. Now he's building a containment platform because agents keep escaping their sandboxes in real incidents. Anthropic. OpenAI. Twice. His own framing draws the line precisely. If containment is achievable through engineering, the risk is manageable. If it isn't, "it's not solvable." That's not confidence. That's a bet the entire industry is currently making in public, live, with no fallback plan stated for the second scenario.