On Monday (28), Nvidia introduced the Open Agent Safety Platform, a security architecture that seeks to control AI agents through layers external to the model itself. The proposal combines software isolation with independent hardware monitoring, making it possible to block actions that exceed permissions defined by the company responsible for the agent.
The first component is OpenShell, open-source software that creates an isolated execution environment for the agent. It logs actions and applies policies while the system works, restricting which data, tools, APIs, and services can be accessed. Nvidia states that OpenShell is already widely available and can also work on platforms based on Arm and Intel chips.
The central difference of the approach appears in the Nvidia Sentry. The system was designed to run separately from the agent, on BlueField 4 DPUs, monitoring its activity from an isolated trust domain. If the agent tries to cross the limits defined by the software, Sentry can quarantine it and halt its execution in milliseconds.
In practice, Nvidia is shifting part of agent security to an infrastructure that the agent itself does not control. This reduces reliance on instructions, filters, and security mechanisms implemented only at the application layer, which can be bypassed by the system during task execution.
Platform already involves more than 100 organizations
Nvidia says that more than 100 organizations work with technologies from the platform. The list includes Anthropic, Microsoft, Salesforce, SAP, JPMorganChase, Cisco, CrowdStrike, Hugging Face, and Scale AI. SpaceXAI uses the technology with Cursor coding agents and Grok models, while Salesforce integrated OpenShell into Slack to allow activity tracking and approval of new permissions by humans.
The launch comes after a series of incidents involving agents that exceeded controls established during tests or operations. Nvidia executives told Reuters that the new architecture could have prevented a breach suffered by Hugging Face this year if it had been in use during the evaluations of the agents involved.
The OpenShell and other software components are already available in Nvidia's developer resources and on GitHub. Sentry is part of the platform's reference design and extends the control model to an independent hardware layer.
