NVIDIA introduced an AI-agent safety platform on September 28, 2026. It combines OpenShell, a runtime with enforceable access rules, and Sentry, a reference design for monitoring outside the agent software. The company presents it as a response to cases in which agents crossed technical boundaries while completing tasks.
OpenShell itself is not new: NVIDIA announced it in March and now says it is broadly available. It is meant to control which files, services and tools an agent can reach while running. Sentry is designed to watch activity on separate BlueField-4 processors and isolate an agent if it attempts to cross its boundary. That moves part of the enforcement away from the agent itself.
NVIDIA names Anthropic, IBM, Salesforce and SAP among the companies involved. IBM describes in its own post how identity and permissions are intended to work with OpenShell. That confirms collaboration, not that every announced integration is already deployed or prevents every bypass.
In coverage of the launch, NVIDIA said the new technology might have stopped agents' earlier access to Hugging Face. That is the vendor's assessment. Neither its announcement nor the independent report provides a test recreating that specific counterfactual and proving the claimed protection.
We record the announcement as a Risk Signal: a concrete industry safety response, not a new incident. Assessing its real protective effect will require published evaluations and operational evidence.