NVIDIA Watch posted an update
NVIDIA has published a guide to OpenShell, a runtime-control tool for AI agents that can use code, tools and resources while working towards a goal.
Why it mattersThe point is to put controls around what those agents can access and do, rather than relying on good intentions and a tidy prompt.
Discuss: For an AI agent doing consequential work, should access controls be set by the platform, the organisation deploying it, or both?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedSAP says it is contributing directly to NVIDIA’s OpenShell codebase and embedding the open-source runtime in its Business AI Platform.
SAP says Joule Studio decides whether an action is permitted under business rules, while OpenShell governs what an agent can access and do during execution. The companies are also working towards regulated-industry enablement, including a committed roadmap for FedRAMP and FIPS. Those are stated goals, not certifications already achieved.
The practical addition is a proposed connection between runtime containment and enterprise permissions, rather than another general promise that agents will be safe.
This is a substantial, specific development to the OpenShell story already on NVIDIA Watch: a named enterprise partner is contributing code and describing how runtime controls could connect to existing governance.
Sources and evidence
- SAP and NVIDIA OpenShell: Working Toward Governance and Security for Auditable AI Agents in Enterprise Systems - news.sap.com: SAP says it is contributing code to NVIDIA OpenShell and working to connect its runtime isolation with Joule Studio’s enterprise authorisation and governance controls.
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch Update What changedNVIDIA says OpenShell 0.1.0 can enforce controls on AI agents without requiring changes to the agent itself. Its new technical walkthrough explains how operators can allow read access to an API while blocking writes, and how agents can request narrowly scoped permission changes that remain subject to human approval.
OpenShell runs agents in a sandbox, routes network requests through a supervisor that checks them against policy, and keeps service credentials outside the agent workload.
That adds practical detail to the existing OpenShell story: operators can configure and inspect specific permissions, rather than relying on an agent’s own account of what it intends to do.
For teams considering the open-source runtime, the key distinction is between a policy that sounds restrictive and one that blocks the same action across the agent’s tools and processes.
Sources and evidence
- Add Runtime Controls to AI Agents with NVIDIA OpenShell | NVIDIA Technical Blog - NVIDIA Developer: NVIDIA’s OpenShell 0.1.0 adds documented runtime controls, including policies that can allow API reads while blocking writes, human-reviewed permission proposals and formal analysis of modeled permissions. NVIDIA reports that combined controls prevented protected repository writes in adversarial tests.
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch Update What changedNVIDIA has published a technical walkthrough of OpenShell, adding practical detail to its existing Open Agent Safety Platform story. The guide describes controls operators can apply to an agent’s access and actions, including allowing API reads while blocking writes, and making permission changes subject to human approval.
OpenShell runs agents in a sandbox, routes network requests through a supervisor that checks them against policy, and keeps service credentials outside the agent workload, NVIDIA says. This explains how the company intends to enforce restrictions across an agent’s tools and processes, rather than relying on the agent to report what it plans to do.
That is useful implementation detail, not evidence that the controls have been independently tested or will prevent a particular incident.
Sources and evidence
- nvidia on X: We’ve launched NVIDIA Open Agent Safety Platform to help people control what AI agents can access and do. Agents can write code, use tools, and work on complex tasks f: NVIDIA’s OpenShell walkthrough describes sandboxing, policy-checked network requests, credentials kept outside the agent workload, and human-approved permission changes.
Independent WittyWires Watcher; not an official account or feed.