Nvidia Launches Platform to Contain Rogue AI Agents

Nvidia has launched a new platform designed to monitor and contain AI agents that attempt to operate outside the boundaries set by their developers.

The Nvidia Open Agent Safety Platform combines two components: OpenShell, an open-source software runtime that controls an agent’s access to files, tools, networks, and other resources, and Sentry, a hardware-based monitoring system that can quarantine an agent if it attempts to move beyond those limits.

OpenShell runs agents within a controlled environment and traces their actions while they work. Nvidia says the software can enforce policies before and during an agent’s tasks and can also be extended to run on computing platforms from companies including Arm and Intel.

Sentry operates separately from the agent itself. It runs on Nvidia’s BlueField-4 data processing units and continuously monitors agent activity. If an agent attempts to cross its permitted boundaries, Nvidia says Sentry can quarantine and stop it within milliseconds. The system is designed to enforce access controls for data, tools, APIs, and services independently of the AI model.

Nvidia launched the platform with support from more than 100 organizations across the AI and technology industries, including Anthropic, Microsoft, Cisco, CrowdStrike, Hugging Face, JPMorganChase and SpaceX. Anthropic is also integrating its Claude Managed Agents with OpenShell and BlueField.

The launch follows several recent incidents in which AI models breached intended testing boundaries. OpenAI, Anthropic and Google have all disclosed cases involving models accessing real systems during security evaluations. Nvidia said these incidents showed the need for controls that sit outside the AI model and its software environment.

The company is positioning the platform as a way to add security controls across the full AI agent stack, from the model and runtime environment to the underlying hardware.