Nvidia’s agent safety platform blocks every action a rule does not allow
Anthropic keeps the agent’s credentials in a vault. OpenShell, from Nvidia, enforces the limits outside the model and logs the decision.
SANTA CLARA — Nvidia announced the Open Agent Safety Platform on 28 September, an open software platform and a reference design for controlling agents from test to deployment. Anthropic is a named collaborator. So are Cisco, CrowdStrike, Microsoft, Palo Alto Networks, Salesforce, SAP, ServiceNow and others on a longer list in the release.
The split of labour is specific. Claude Managed Agents holds the credentials an agent needs in a vault, so the agent does not see them. The agent loop runs on a separate server from the sandbox where the work happens. Nvidia OpenShell, which is open source under Apache 2.0, is meant to govern what the agent can run and reach. Anthropic said it blocks everything unless a rule allows it, checks each tool, and logs every allow and every block. The rules sit outside the model.
That is a different offer from a prompt that says “do not do that.” A prompt is inside the system the agent can argue with. A rule enforced outside it is something an auditor can read. The buyer should ask to see one blocked action and the log line for it before treating the platform as control.
54east’s operations-agent brief uses the same test. https://54east.ai/operations-agents/
