Nvidia has announced the free, open-source Nvidia Open Agent Safety Platform, a set of controls designed to govern autonomous AI agents across the agent runtime, computing resources and underlying hardware. SiliconANGLE reports that the platform combines Nvidia’s OpenShell software with Nvidia Sentry.
Controls beyond the model
The platform reflects Nvidia’s argument that prompts and application-level restrictions are not enough to contain agents that can act independently. Its approach places enforcement across three layers: the agents themselves, the compute infrastructure running them and the hardware supporting that infrastructure.
OpenShell provides the runtime environment. Nvidia describes it as an open-source system for operating groups of autonomous agents in isolated sandboxes, while tracing their actions and applying policies outside the guardrails of the large language model they use. The software is optimized for Nvidia Vera central processing units and is also compatible with Intel and Arm CPUs.
Hardware-level monitoring
Nvidia Sentry is the platform’s hardware-oriented component. It is a reference system design based on Nvidia DOCA and is intended to run on BlueField-4 data processing units. Nvidia says Sentry can inspect agent requests, verify identities and actions, and stop an agent when it exceeds its assigned boundaries.
That design moves part of agent oversight below the application layer, allowing controls to cover the infrastructure that executes an agent rather than relying only on the model or software wrapping it. The source describes Sentry as a reference design, so its stated capabilities remain Nvidia’s characterization of the system.
Reported users and limits
According to Nvidia, SpaceXAI is using the platform with Cursor coding agents and the Grok large language model. Salesforce has integrated OpenShell with Slack for human approvals, visibility and audit tracking, while SAP is using it with its SAP Business AI Platform. Nvidia also names Figure AI, Gecko Robotics and Skild AI as OpenShell users in robotics systems.
Nvidia says the platform is available through its developer resources and GitHub. The report does not provide independent testing or evidence showing how reliably the controls perform at large scale, leaving real-world effectiveness unresolved.
AI-generated editorial illustration; not a documentary photograph.
Source: Nvidia debuts enhanced safety controls to rein in rogue AI agents
SIGNAL TITANS
Leave a Reply