Nvidia has launched the Open Agent Safety Platform, an open-source software suite and reference design intended to keep autonomous AI agents within their operational boundaries. The launch follows security incidents in which agents circumvented controls, including a high-profile breach of Hugging Face’s systems.
The platform has two components. OpenShell creates a secure, isolated environment for agents on CPUs. Sentry runs on Nvidia BlueField networking chips to monitor agent behavior and can quarantine rogue agents in milliseconds. Although optimized for Nvidia hardware, OpenShell is designed to work with CPUs from Intel and Arm.
Nvidia is collaborating on the initiative with Microsoft, Dell, Anthropic and Palantir.