NVIDIA OpenShell and Sentry: where INTYGA adds human authorization
Keep people in control of production changes: INTYGA binds human approval to the exact action your service executes. This post proposes how that composes with OpenShell's runtime controls and Sentry; it is not a shipped or tested integration.
Give agents room to prepare work while keeping people in control of the changes that matter. INTYGA lets your application require a human-signed approval for an exact production release, payment or permission change. The executing service verifies that approval before acting. For security and platform teams, this makes the human decision an enforceable condition of execution.
On September 28, NVIDIA announced its Open Agent Safety Platform: OpenShell runtime software together with a reference system design featuring NVIDIA Sentry on BlueField-4 data processing units. The announcement puts controls outside the agent at the center of the architecture. An agent can be useful while its execution environment assumes it may behave unexpectedly.
OpenShell itself was announced in March alongside NemoClaw. September's news brings it into a broader safety platform with an independent hardware enforcement layer. For teams connecting agents to consequential systems, the useful question is what each layer checks before work proceeds.
OpenShell controls the environment in which the agent works
NVIDIA's runtime walkthrough describes sandbox isolation, network and API restrictions, and provider credentials substituted at the boundary, keeping the real credentials outside the agent workload. An operator can control the destinations and operations available to the agent while keeping the underlying service credential out of its environment. Policy proposals wait for human review by default; the agent cannot approve its own proposal.
OpenShell's policy prover checks that a policy grants no more access than a boundary policy the operator defines. The check covers filesystem access, process identity, Landlock settings, network connections and REST requests; a policy that relies on anything else, such as MCP rules, is reported as unsupported, and a check that cannot finish is inconclusive. Only a within-boundary result passes, and NVIDIA's documentation is explicit that even a pass does not mean the policy is safe for a particular task. A proof about those permissions should be read with that scope, rather than as proof that every outcome of a permitted task is desirable.
Sentry places another control outside the host
In NVIDIA's Sentry reference architecture, BlueField-4 tracks and enforces agent activity in separate hardware, which NVIDIA's announcement calls an isolated, out-of-band trust domain. NVIDIA DOCA connects this layer to OpenShell policy and helps correlate agent activity, access and policy decisions. The design gives the infrastructure a place to observe and intervene even when the agent's host cannot be trusted. Sentry is described as an optional layer alongside OpenShell; the documented hardware design should not be confused with a requirement that every OpenShell user own a BlueField system.
That separation is valuable. A control outside the workload can refuse an operation regardless of the agent's explanation for requesting it. INTYGA applies that principle at the service performing the consequential action: your application checks a human signature over the exact request before it acts.
Use INTYGA to approve the release that actually reaches production
Consider an agent preparing a software release. It reads source, builds an image and runs tests inside a restricted environment. Runtime policy limits the repositories, files and services it can use. An operator may approve an additional endpoint needed to complete the build. That approval changes the agent's available access.
The release owner still needs to decide whether a particular artifact should enter production. Suppose the request names image digest A, service billing-api and environment production. The deploy service can require a human signature binding those values, verify them against the request it will execute, and reject a request for digest B or a different service. Naming an immutable digest also prevents a mutable image tag from changing what an approval refers to.
The distinction is the scope and lifetime of the authorization. Permission to use a deployment API can cover many future calls. A signed release approval can cover one specific request, with an expiry and a nonce the executing service consumes once. Both decisions can be useful in the same workflow.
Agent prepares artifact within runtime policy
|
v
Release request: digest A + billing-api + production
|
human signs request
|
v
Deploy service verifies its own intended action
- trusted approver keys and required quorum
- exact target, action and signed parameters
- approval ceremony, requester and expiry
- durable single-use nonce reservation
|
v
Deploy the approved digest using a service-held credentialWhat INTYGA gives your team
INTYGA gives the release owner a concrete decision: approve this artifact, for this service, in this environment. They sign with a passkey in the browser. Your executing service verifies the receipt locally against approver keys it already trusts and the action it is actually about to perform. A changed signed parameter fails verification. Your team can preserve the approval as evidence that can be checked independently.
- Focus human review on consequential actions: choose the operations that require approval while agents prepare the work within your existing controls.
- Keep approval keys with the approver: INTYGA never holds the private key that signs the human's authorization.
- Verify on your terms: your service checks the receipt using its own expected parameters and trusted keys, without calling INTYGA for receipt verification.
The integration must also prevent replay and bypass. The service reserves the nonce durably before execution, atomically with a database change where possible. For an external deployment call, a failed attempt can consume the approval without completing the release. Only the enforcing service should hold the deployment credential, and the agent must have no alternative path to the same production change.
For INTYGA's AI-agent receipts, the executor also owns the expected configuration and durable session state, including the session head and aggregate budget. Those checks must use trusted runtime state. A digest supplied by the agent does not establish what it actually ran.
The service-side integration walkthrough covers these verification and replay responsibilities. This is a proposed composition of controls, not a shipped or tested INTYGA integration with NVIDIA OpenShell or Sentry.
Keep evidence of the human decision
A runtime policy decision records whether configured controls permitted a request. A human-signed receipt lets a relying party check which exact request an approving credential signed. A downstream execution record is needed to establish whether the operation actually completed. Keep those records linked, and keep their meanings distinct.
Human authorization has limits too. A person can approve a bad release, and a correctly signed digest does not prove that the image is free of malicious code. Compromise of the service performing the verification remains outside the receipt's protection. Runtime isolation, credential controls, artifact review and exact-action authorization therefore have continuing jobs to do together.
Start with one production release. Connect INTYGA at the deploy service, keep the deployment credential there, and require a receipt bound to the artifact digest, service and environment. Demonstrate that a changed request and a reused approval are refused. That gives your team a concrete control to evaluate before expanding to more workflows.
Put INTYGA in front of your next high-impact action
Put INTYGA to work on one workflow your team needs to control. Talk to us about INTYGA to map the approval and execution boundary, or follow the integration quickstart to build and test your first approval flow.