英伟达发布 Agent 安全栈 OpenShell 与 Sentry
英伟达推出把控制权放在 Agent 外部的安全方案：OpenShell 限制其可访问范围，Sentry 独立监控行为并在越界时关停。
We spent the last two years asking how much autonomy we can give AI agents. I think the industry is finally asking the better question: how do we stop them when they’re wrong? $NVDA just launched an agent safety stack that puts controls outside the agent itself. OpenShell restricts what an agent can access, while Sentry can independently monitor its behavior and shut it down if it crosses the boundary. That architecture matters more than another benchmark. We don’t secure banks by asking employees to promise they won’t access the wrong account. We build permissions, audit trails and controls around them. AI agents will need the same thing. The future isn’t unlimited autonomy. It’s useful autonomy with accountable boundaries.