Nvidia Cracks Down on Rogue AI Agents

Nvidia has introduced two tools, OpenShell and Sentry, to contain AI agents and prevent them from running wild. The move comes after a hack disclosed this summer by Hugging Face highlighted the need for better agent security.

AI agents can take multi-step actions on their own, making them useful but also harder to control once they're plugged into real company systems. Nvidia's approach shifts the focus from 'trust the model to behave' to 'constrain what it can do'. OpenShell uses hardware protections on Nvidia central processor chips to isolate an agent in a restricted environment, while Sentry adds a separate Nvidia chip that can cut the agent off if it tries to escape.

Nvidia is working with Arm Holdings and Intel so the same idea can run on their processors. The company is launching with dozens of partners, including Anthropic. By baking containment into chips and pairing it with an independent 'off switch', Nvidia is positioning agent safety as something buyers can verify, a feature that ends up in security and compliance checklists.