Extending Admin Control to the Runtime Layer with NVIDIA OpenShell and Slack

As AI agents take on longer-running work across enterprise systems, they need broader access to tools, data, credentials, and services to get that work done. The challenge is giving agents enough autonomy to be useful while making sure humans stay in the loop on what those agents can access and do.

The NVIDIA Open Agent Safety Platform is an open source platform and reference architecture for AI agent security and governance that includes the NVIDIA OpenShell secure runtime. OpenShell allows organizations to define and enforce what an AI agent can access and do. When an agent hits one of those boundaries, that’s where Slack comes in. With the Slack Admin Bridge for OpenShell, those access exceptions flow natively into Slack, bringing approval decisions to admins right where they already work. Overall agent activity also flows into Slack so admins can maintain real time analytics and health of agents. 

OpenShell provides a secure runtime. Slack Admin Bridge provides the human control surface.

Together with NVIDIA, we built a Slack Admin Bridge for OpenShell that brings this incremental human oversight of agents directly into Slack.

When an agent running in OpenShell needs to connect to something outside its existing network policy, the request surfaces in Slack as an interactive approval card. Admins can review requests without leaving their workflow, and the conversation, debate, and audit trail around those decisions live in one visible place instead of scattered across tools. Real-time analytics in Slack’s App Home give admins a running view of outstanding requests, past decisions, and agent activity from any device, and the bridge runs in your own secure environment, connecting to Slack with permissions scoped down to sending information and ingesting approvals only, so no message, channel, or user data ever flows out of Slack.

The integration is designed with enterprise security in mind. OpenShell credentials remain outside Slack, and the bridge is scoped to only send audit events and ingest approvals.. The result is a clear separation between the agent requesting access, the runtime enforcing the policy, and the human authorizing the exception.

Giving agents autonomy without giving up control

For customers, the goal is simple: let agents do more while keeping admins in control. OpenShell enforces incremental controls at the runtime, and Slack brings the resulting exceptions to the right person in real time, with the discussion, approval, and audit trail all visible in one place. Because the integration runs in your own secure environment with permissions scoped to only send information and ingest approvals, teams get real-time control and full auditability without any message, channel, or user data ever leaving Slack.

Looking ahead

Salesforce and NVIDIA are also exploring how OpenShell could integrate with MuleSoft Agent Fabric, pairing Agent Fabric’s governance and orchestration through Omni Gateway with OpenShell’s secure runtime capabilities. Together, the technologies could help enterprises govern which agents can access systems while enforcing what those agents can do at runtime.

Get Started

The OpenShell + Slack integration is available for developers to use today. Once OpenShell is set up, developers can connect the Slack Admin Bridge and choose which admins and channels should receive agent approval requests. For setup instructions and more details, see here. 

For more information on the NVIDIA Open Agent Safety Platform and the technologies announced today, see NVIDIA’s announcement here.