Nvidia on Sept. 28 introduced the Open Agent Safety Platform, an open platform designed to handle security for AI agents from testing through deployment.

Nvidia said it consists of the open-source software OpenShell and the hardware-based reference design Sentry. OpenShell is runtime software that sets a security boundary in the environment where an agent runs, tracks all actions and enforces policies. It minimises performance degradation on Nvidia's agentic AI CPU, Vera. Because it is open-source, it can be expanded to third-party computing platforms such as Arm and Intel.

Sentry is an external monitoring device that runs on the BlueField-4 DPU. If an agent tries to break out of the software boundary, it isolates and blocks it within milliseconds. Nvidia explained that it operates in an independent trusted domain that neither the agent nor an attacker can see. Nvidia said a common factor identified in recent security incidents was that agents bypassed application-layer security controls to complete tasks.

Nvidia CEO Jensen Huang (젠슨 황) said, "AI’s potential to benefit society can be realised only by solving safety issues," adding, "Safety and security require full-stack engineering."

Participation by companies has also expanded sharply. They include Anthropic, Cisco, CrowdStrike, Dell, Figure, HPE, Hugging Face, JPMorgan Chase, Microsoft, Palantir, Palo Alto Networks, Perplexity, Red Hat, Salesforce, SAP, Scale AI, ServiceNow and SpaceXAI. More than 100 companies are using the related technology.

OpenShell and related skills can be downloaded from Nvidia's developer page and GitHub. It is also linked to the Open Secure AI Alliance, which Nvidia launched with about 120 institutions. The alliance is operated by the Linux Foundation.