#

Key takeaways

• NVIDIA announced the Open Agent Safety Platform on September 28, 2026.

• The platform combines OpenShell and Sentry in one layered control design.

• NVIDIA says OpenShell traces agent actions and enforces policy.

• NVIDIA says Sentry can monitor behavior and quarantine agents in milliseconds.

• The release names more than 100 organizations joining or using the platform at launch.

NVIDIA's new platform for agent safety

NVIDIA announced the Open Agent Safety Platform on September 28, 2026. The company describes it as an open software platform and reference system design. Its goal is to add governance and control from agent testing through deployment.

The announcement brings together two parts. OpenShell is described as open-source secure runtime software. Sentry is described as an out-of-band watchdog reference design for monitoring behavior.

NVIDIA frames the system as layered. It spans software and hardware. The company also says organizations can deploy elements according to their requirements.

What OpenShell and Sentry do

According to NVIDIA, OpenShell traces agent actions and enforces policy. The company says it is broadly available. It also says OpenShell can be extended to work with third-party compute platforms, including Arm and Intel.

Sentry is the hardware-side part of the announcement. NVIDIA says it runs on NVIDIA BlueField-4 DPUs. The company says it can quarantine agents that attempt to cross defined boundaries in milliseconds.

These descriptions come from NVIDIA's own announcement. They explain the intended design of the platform. They do not independently prove how well the controls work in real deployments.

Ecosystem participation at launch

NVIDIA says the launch includes participation from across the AI ecosystem. The release names Anthropic, Cisco, CrowdStrike, Dell, Hugging Face, JPMorganChase, Microsoft, Perplexity, Red Hat, Salesforce and ServiceNow.

The company also says over 100 organizations are joining or using the platform at launch. That statement should be read as NVIDIA's own description of the relationships involved. The release does not provide a separate independent count or verification.

A separate technical blog published the same day discusses continuous monitoring. NVIDIA says it also lists five design principles. The source material does not provide those principles in full, so this article does not add them.

Why the announcement matters

The release positions agent safety as a lifecycle issue. It starts at testing and continues through deployment. That framing suggests NVIDIA wants governance built into the system, not added later.

The layered design is also notable. NVIDIA combines software tracing, policy enforcement, and hardware monitoring in one announcement. The company says organizations can choose the elements they need.

The source material does not claim the platform prevents incidents in all cases. It also does not provide performance data. So the safest reading is that NVIDIA is presenting a control framework, not a proven guarantee.

Operational considerations

The announcement points to a few practical considerations. First, the platform depends on how organizations configure and deploy its parts. Second, NVIDIA says the system can work with third-party compute platforms for OpenShell, which may matter for mixed environments.

Third, the release emphasizes continuous monitoring. That suggests ongoing oversight is part of the design. Fourth, the quarantine claim for Sentry is specific to NVIDIA's description and should be treated as such.

Morocco relevance

The source reports no Morocco-specific deployment, partnership, or market detail. For readers, the general lesson is that agent safety tools are being designed as layered systems. Any local evaluation should still depend on the platform's own technical fit and governance needs.

Bottom line

NVIDIA's Open Agent Safety Platform is an announcement about control, monitoring, and policy enforcement for AI agents. It combines OpenShell and Sentry into one architecture. The release also signals broad ecosystem interest, while leaving real-world effectiveness to be proven over time.