NVIDIA announced the Open Agent Safety Platform on 28 September 2026, describing an open software platform and reference system design for securing AI agents from testing through deployment. The release centers on OpenShell, open source runtime software for agent boundaries, and NVIDIA Sentry, a reference design for out-of-band monitoring and enforcement on BlueField-4 DPUs.
This is a material AI infrastructure story because it moves agent governance below the chat or application layer. NVIDIA's argument is that long-running agents need boundaries enforced outside the model and agent harness, especially when agents can access tools, data, APIs, code, robotics systems or enterprise workflows.
What NVIDIA announced
NVIDIA says OpenShell provides a secure runtime boundary that traces agent actions and enforces policy as agents run on NVIDIA Vera CPUs. The company says OpenShell is now broadly available and can be extended to third-party compute platforms, including Arm and Intel.
Sentry is different. It is a reference system design that runs on NVIDIA BlueField-4 DPUs, monitors agent behavior independently and can quarantine agents that attempt to move outside their boundaries. NVIDIA describes this as in-silicon enforcement from an isolated trust domain.
The announcement also names a large partner set, including Anthropic, HPE, Microsoft, Salesforce, SAP, Scale AI, Palo Alto Networks and others. HPE separately says OpenShell is now generally available and that HPE Private Cloud AI is integrating it, with HPE-specific availability planned for Q4 2026. That distinction matters: OpenShell availability is current, while some enterprise integrations remain planned or deployment-dependent.
Why this matters
Recent agent incidents have made one point hard to ignore. A prompt can say "stay in scope," but the agent still needs an environment that constrains file access, network access, tool calls, identity, execution and escalation. The enforcement boundary should not depend only on the same model that is deciding what to do.
For governance and compliance teams, this changes the evidence question. A runtime boundary can produce records that show what the agent tried, which policy applied, which exception was requested, what a human approved and what the final effect was. That is more useful than a policy document that cannot prove whether it held at action time.
Maetra's runtime guardrails comparison makes the same distinction. Policy review is necessary, but agent work also needs action-time enforcement and evidence.
What buyers should test
The launch is promising, but organizations should validate the exact deployment path. OpenShell should be tested against the agent frameworks, tool calls, network destinations, file systems and identity providers actually in use. Teams should also confirm how audit events are exported, how exception approvals are bound to exact actions, and what happens when a model asks for more permission.
For Sentry and hardware enforcement, buyers should verify which hardware is required, whether the reference design is available for their environment and how monitoring behaves when agents run across cloud, on-premises, robotics or third-party compute.
What remains uncertain
The public announcement does not independently prove effectiveness, false-positive rates, quarantine reliability or customer outcomes. Partner quotes show ecosystem interest, not uniform production deployment. Some integrations and hardware-based controls depend on product lead times or partner release schedules.
Maetra analysis
NVIDIA is pointing at the right layer. The next generation of agent safety will not be only better prompts or better model refusals. It will be enforceable task boundaries, runtime policies, identities, human approval paths and evidence that survives after the action.
The strongest governance pattern is layered. Inventory the agents and capabilities. Define what each agent may read, write, execute and call. Route exceptions to a human only when policy requires it. Preserve the policy decision, runtime trace and final effect. Then test the boundary with failure cases, not only happy paths.
If OpenShell and similar systems become common, agent security reviews will become more concrete. The question will shift from "Did we tell the model not to do that?" to "Which boundary enforced the rule, what evidence did it produce and what changed after the exception request?"
Sources
Primary source: NVIDIA Open Agent Safety Platform announcement.
Corroboration: HPE on NVIDIA governed agentic AI integration.