NVIDIA launched the Open Agent Safety Platform on Monday, a set of software and hardware controls meant to keep AI agents from breaching the limits set for them.
The platform has two parts. OpenShell is open-source runtime software that traces an agent’s actions and enforces policy as it runs, and it is available now on GitHub and NVIDIA’s developer site. Sentry is a watchdog that runs on NVIDIA BlueField-4 data processing units, separate from the machine hosting the agent.
NVIDIA says Sentry can quarantine an agent that tries to move outside its boundaries in milliseconds. The company describes it as “invisible to agents and attackers,” which is the point of putting it on separate hardware.
OpenShell is tuned for NVIDIA’s Vera CPUs, but NVIDIA says it can be extended to chips from Arm and Intel. Sentry is described as a reference system design, and NVIDIA’s availability notes list only OpenShell and its skills for download. NVIDIA introduced OpenShell in March; the new part is the bundle and the wider partner list.
The launch follows a series of incidents in which AI agents got past their test environments. In July, OpenAI said two of its models escaped a sandbox and breached Hugging Face’s production systems to obtain answers to a cybersecurity benchmark.
Justin Boitano, NVIDIA’s vice president of enterprise AI, told reporters the platform could have stopped that breach if frontier labs had used it during model evaluation, the Associated Press reported.
NVIDIA says more than 100 organizations are working with the platform’s technologies. Named partners include Anthropic, Microsoft, SpaceXAI, JPMorganChase, Salesforce, SAP, and Hugging Face.
NVIDIA agreed on September 2 to acquire Hugging Face, and the deal is expected to close in the first half of 2027, according to an SEC filing.
Some of the integrations are concrete. Salesforce has connected OpenShell to Slack so teams can view agent activity and approve or reject requests for extra permissions. SpaceXAI says it is using the platform for Cursor coding agents and Grok models.
Anthropic said NVIDIA’s platform adds another layer of control to Claude Managed Agents, which already runs an agent’s decision-making on a separate server from the sandboxes where its work executes.