← Back to news

Cyphrex Intelligence

Nvidia Put Agent Safety Into Open Software and Silicon on the Same Day

September 28, 2026

Category: Security | Reading time: 6 min

The Launch

On 28 September 2026, Nvidia announced the Open Agent Safety Platform, a reference design that spans software and hardware for governing AI agents from testing through deployment.

The platform has two layers. OpenShell is open-source runtime software that sandboxes agents and enforces policy outside the model and the agent harness. Sentry is an out-of-band watchdog that runs on Nvidia BlueField-4 data processing units and can quarantine an agent that steps outside its boundary in milliseconds.

Nvidia said organizations can deploy the layers independently or together. OpenShell is broadly available at version 0.1.0. Sentry ships as part of the reference system design for BlueField-4.

What OpenShell Actually Enforces

OpenShell does not rewrite the agent. It defines which systems and data the agent may touch, then enforces those permissions in the runtime.

SecurityWeek reported three components: a gateway that manages sandbox lifecycles and policies, a sandbox that applies kernel-level controls to filesystem and process activity, and a supervisor paired with each sandbox that checks outbound requests against policy.

The supervisor can inspect configured HTTP, GraphQL, and Model Context Protocol traffic. That means a read through an API can be allowed while a write on the same endpoint is blocked. Credentials stay at the gateway and are injected at runtime rather than baked into the workload. Policy decisions are recorded in an OpenTelemetry Schema Framework audit trail.

OpenShell was introduced in March 2026 and now supports agents including Codex, Claude Code, Pi, and Hermes. As open source, Nvidia says it can extend beyond Nvidia Vera CPUs to third-party platforms, including Arm and Intel.

What Sentry Adds

Sentry runs on BlueField-4, which sits on the node's path to the model and outside the agent's host. Nvidia's claim is that the DPU can observe agent behavior and enforce policy even if the host is compromised.

Built on Nvidia DOCA, Sentry is described as providing attested telemetry, agent identity verification, and zero-trust access policy for data, tools, APIs, and services. When an agent attempts to move outside its software boundary, Sentry quarantines and stops it in milliseconds.

In an Nvidia Vera Rubin POD, each compute tray includes a BlueField-4. Organizations already on Vera with BlueField-4 can turn the protections on with a software update. Nvidia says the platform is also compatible with other hardware.

Who Showed Up

Coverage of the launch listed more than 100 organizations in the partner and early-adopter orbit, including Anthropic, Salesforce, SAP, CrowdStrike, Palo Alto Networks, Cisco, and SpaceXAI for Cursor and Grok workloads. OpenAI did not appear on the public partner list released with the announcement.

That absence is a market signal, not a technical verdict. It does show that a major chip vendor is organizing the agent-safety conversation around runtime boundaries and hardware enforcement, and that a large slice of the ecosystem is willing to stand next to that framing on day one.

What This Changes for Deployers

Enterprise agent security has spent two years debating prompts, model refusals, and post-hoc logs. Nvidia's launch treats the decisive layer as the execution boundary: what the agent may touch, whether that boundary holds when the host is hostile, and whether a stop decision can fire before irreversible actions complete.

OpenShell and Sentry are infrastructure answers to that problem. They are not a substitute for verified agent identity, signed write-time records, or independently auditable evidence packs that a compliance team can hand an examiner. They do raise the floor for anyone running agents on Nvidia hardware and make kernel- and silicon-enforced boundaries a procurement question, not a research topic.

Teams already shipping production agents should treat this as a category shift: runtime quarantine and hardware watchdogs are now on the vendor roadmap of the company that sells the GPUs those agents often sit next to.

Source: SecurityWeek · NVIDIA Technical Blog · GlobeNewswire

Message the founder