NVIDIA lance the Open Agent Safety Platform to monitor and contain artificial intelligence agents capable of acting autonomously. According to Nvidia, its Sentry hardware component can quarantine an AI agent that leaves its security perimeter within a few milliseconds.
The Open Agent Safety Platform combines two elements. OpenShell, available as open source, creates a secure execution environment and defines the resources that an AI agent can access, including files, networks, handles, APIs, and the various tools necessary for its mission.
Nvidia adds Sentry as a separate monitoring layer. This firewall runs on BlueField-4 DPU infrastructure processors, separate from the environment in which the agent executes its tasks, so it can intervene even when the agent attempts to bypass software protections.
The system monitors agent requests and responses, verifies their identity, and applies access policies to data, tools, APIs, and services. If an agent attempts to cross boundaries set by the administrator, Sentry can isolate it and shut it down within milliseconds, according to Nvidia.
This architecture responds directly to a weakness in the protections built into AI applications: an agent seeking to circumvent its own restrictions cannot be the sole mechanism responsible for monitoring its behavior. Nvidia therefore places some control outside of the agent and its software environment.
More than 100 actors join the project
Nvidia presents its platform as an open project intended to cover the entire technical chain, from software to processors and robotic systems. OpenShell runs on Nvidia Vera processors, but its code can also be ported to Arm and Intel platforms.
More than 100 organizations already work with the platform’s technologies. Anthropic, Microsoft, Cisco, CrowdStrike, Dell, HPE, Hugging Face, IBM, JPMorganChase, Palo Alto Networks, Perplexity, Salesforce, SAP, Scale AI, ServiceNow and SpaceXAI are among the companies cited by Nvidia.
Anthropic is notably working with Nvidia to integrate OpenShell and BlueField into its managed Claude agents. Salesforce is also working to integrate OpenShell into its tools, while Scale AI uses the reference design for its business and government systems.
Open Agent Safety Platform software components are now available through developer resources from Nvidia and GitHub. Sentry, for its part, remains a reference design built around the BlueField-4 DPU and not a standalone product marketed in this way.
