NVIDIA Announces Open Agent Safety Platform to Contain AI Agents
The pitch turns agent safety from a policy argument into a hardware sale: OpenShell runs on Vera CPUs and Sentry runs on BlueField-4 DPUs, so containment requires buying NVIDIA silicon at both the compute and the monitoring layer.
Key Facts
- NVIDIA announced the Open Agent Safety Platform, a framework for monitoring and constraining AI agents, on September 29, 2026.
- NVIDIA Sentry can stop an AI agent attempting to move outside its boundary in milliseconds, according to NVIDIA.
- NVIDIA named Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, ARM, and Intel as partners on the Open Agent Safety Platform.
- Anthropic said its Claude Managed Agents API suite can use NVIDIA OpenShell, with Rakuten and Notion already using it.
Reporting from 2 sources: GameBusiness.jp, GIGAZINE.
NVIDIA announced the Open Agent Safety Platform, a framework for monitoring and constraining AI agents that act autonomously. The platform has two components. NVIDIA OpenShell is open-source software that runs on an NVIDIA Vera CPU and checks each tool an agent tries to use, letting operators set rules on files and network connections the agent can reach and blocking anything outside those rules. NVIDIA Sentry is a separate monitoring system that runs on the BlueField-4 data processing unit and can stop an agent that moves outside its boundary in milliseconds. NVIDIA CEO Jensen Huang called the platform a browser for agents and said trust in how AI is built and deployed is a condition for the AI industry's success. A company spokesperson said the July incident in which an OpenAI AI agent intruded into Hugging Face could possibly have been prevented with the platform.
NVIDIA lists five basic principles its agent systems are built to follow. Policies must be verifiable, so a prover shows before an agent runs that the policy cannot escape the operator's intent. Monitoring should be out-of-band, meaning the control system cannot live inside the agent or within its reach, and the agent does not need to know it is being watched.
Controlling the path to the model gives the operator an observation point and a kill switch to interrupt it. The agent's authority extends through functionality that can verify its thought process. Research institutes, companies, and hardware providers each own a layer, and the runtime and its policy language must be open so any provider can plug in.
Partners already named include Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, ARM, and Intel. Anthropic is also working with NVIDIA to integrate cloud-managed agents into OpenShell through its Claude Managed Agents API suite, which companies such as Rakuten and Notion are already using.
Huang framed the stakes plainly: "We must build not only the most capable AI, but the most trusted AI."
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.