AI agents can run code, access applications and call external tools. As they gain autonomy, enforcing limits becomes more important, especially when an instruction is misunderstood or a component is compromised. On September 28, 2026, NVIDIA announced its Open Agent Safety Platform, an open software platform and reference design intended to move agent controls beyond model instructions.
The announcement brings together two distinct components: OpenShell, an open-source runtime that enforces policies and isolation, and Sentry, a reference design for independent monitoring on NVIDIA BlueField-4 data processing units. The central idea is to enforce some restrictions outside the agent's own process, where an incorrect model response should not be able to override them.
OpenShell: limits enforced at runtime
OpenShell runs agents in isolated environments and defines which files, services, networks and tools they may reach. According to NVIDIA's technical article, its architecture uses Gateway, Supervisor and Sandbox components to manage agent workloads, inspect outbound requests and apply filesystem and process controls. Explicit permissions replace reliance on prompts that merely instruct a model not to access certain systems.
NVIDIA says OpenShell is now broadly available as open-source software. It highlights optimizations for its Vera CPU while stating that OpenShell can also be extended to third-party compute platforms, including Arm and Intel. Extensibility does not imply identical integration or performance across hardware configurations.
Sentry: an observer outside the agent
Sentry is described as a reference design that uses BlueField-4 DPUs to monitor agent activity from a separate trust domain. NVIDIA says the system can verify identities, inspect requests and responses, enforce access policies and quarantine agents that cross defined boundaries in milliseconds. The timing is a vendor-described capability, not a performance guarantee for every deployment.
That separation matters. If an agent or part of its execution environment fails, an external layer may provide additional oversight. It does not remove the need to configure policies carefully, minimize privileges and audit outcomes.
Infographic: how the two layers work together
What is available now, and what still needs evaluation?
The public OpenShell repository allows developers to inspect the source, license and installation instructions. Its documentation covers policies, sandboxes and network controls. Sentry, meanwhile, was announced as part of the platform's reference design. Teams considering it still need to establish the availability, infrastructure requirements and integration status relevant to their own environment.
NVIDIA says more than 100 organizations are working with technologies associated with the platform. Its announcement names Anthropic, Microsoft, SAP and Salesforce among others, but collaboration, announced integration and proven production deployment are not interchangeable.
Why should developers care?
Consider an agent authorized to read a repository and open a pull request, but not to extract credentials or perform irreversible production operations. The model can continue receiving goals in natural language; access limits should still be checked by infrastructure controls that do not depend on the agent choosing to follow instructions.
Teams evaluating this approach can begin by inventorying accessible resources, establishing least-privilege policies and testing attempts to cross defined boundaries. They should also assess secret handling, outbound calls, audit trails and human approval for sensitive actions.
The Open Agent Safety Platform is a layered-security proposal for autonomous AI systems. Its practical value will depend on each organization's integrations, testing and ongoing operations. The announcement expands the available options, but it does not replace independent risk assessment for a specific application.
Official sources
- NVIDIA Newsroom: Open Agent Safety Platform announcement, September 28, 2026.
- NVIDIA Developer: runtime controls with OpenShell, September 28, 2026.
- NVIDIA/OpenShell on GitHub, project code, license and documentation.
- Official OpenShell documentation, concepts, policies and deployment.