Image Credit: Shutterstock
- Nvidia has launched the Open Agent Safety Platform to keep autonomous AI agents inside defined operating boundaries.
- The platform combines the open-source OpenShell runtime with a BlueField-4-based Sentry monitoring layer that can independently quarantine agents.
- More than 100 organizations are working with Nvidia’s safety stack, including Anthropic, Microsoft, JPMorgan Chase, and SpaceXAI.
Nvidia has launched the Open Agent Safety Platform, a software, and hardware security stack designed to keep autonomous AI agents inside defined operating boundaries.
The platform combines Nvidia’s open-source OpenShell runtime with Sentry, a BlueField-4-based monitoring layer that can independently quarantine agents when they attempt to operate outside approved limits.
OpenShell Restricts File, Network, and API Access
OpenShell runs AI agents inside isolated environments and applies policies to file, network and process access. Permissions are denied by default and granted where required, while allow and deny decisions can be logged for auditing.
The runtime sits outside the agent workload, allowing security teams to restrict API calls and protect credentials without rewriting the agent itself. OpenShell supports agents including Codex and Claude Code and can run across local, cloud and on-premises infrastructure.
The approach is designed for organizations giving autonomous agents access to internal files, credentials, APIs, and other systems while maintaining controls outside the underlying AI model.
BlueField-4 Sentry Can Quarantine Agents in Milliseconds
Nvidia Sentry adds a separate hardware enforcement layer through BlueField-4 data processing units.
The system monitors agent activity independently of the CPU or GPU running the workload and can quarantine an agent in milliseconds if it crosses its defined software boundary.
Sentry uses Nvidia’s DOCA software to monitor requests and enforce access policies for data, tools, APIs, and services independently of the host system.
More Than 100 Organizations are Working With Nvidia’s Agent Safety Stack
Nvidia says more than 100 organizations are working with Open Agent Safety Platform technologies, including Anthropic, Microsoft, JPMorgan Chase, Salesforce, SAP, Scale AI, ServiceNow and SpaceXAI.
Salesforce has integrated OpenShell with Slack, so teams can review agent activity and approve additional permissions, while SAP is embedding the runtime into Joule Studio.
OpenShell and other platform software are available through Nvidia’s developer resources and GitHub. Sentry is included in Nvidia’s BlueField-4 reference system design, giving organizations a separate hardware layer for enforcing controls around autonomous AI agents.
