NVIDIA's new safety platform runs the controls that stop an AI agent on separate software and hardware, out of the agent's reach. The partner list shows which of your vendors to ask next.

Every incident NVIDIA describes in announcing its Open Agent Safety Platform follows one pattern: the agent got around the security controls built into its application to finish the task it was given. The platform, announced September 28, puts its controls outside the application in two parts.

The first part is OpenShell, open source software that records every action an agent takes and enforces a set of rules while the agent runs on NVIDIA Vera processors. The second is Sentry, a blueprint for a watchdog that runs on separate NVIDIA BlueField-4 chips, watches agent behavior, and stops an agent that steps outside its limits. NVIDIA says Sentry acts in milliseconds (NVIDIA, 2026).

Controls on separate hardware sit where an agent cannot reach them

NVIDIA says Sentry is invisible to agents and attackers because it runs on its own chip, apart from the software the agent uses. Infrastructure teams have kept their management systems separate from the workloads they manage for years, and Sentry applies the same separation to agent behavior. A control that lives inside the agent's own software shares that software's weaknesses, so moving it to a separate chip changes the risk.

In July I described an agent as a model wrapped in identity checks, permissions, and other controls. Sentry covers identity and permissions by checking who an agent is and enforcing what it may access from separate hardware.

NVIDIA's release says Anthropic's Claude Managed Agents run the agent's decision loop on one server and its tasks in isolated workspaces, called sandboxes, on another. The release adds that integrations with OpenShell and BlueField let companies enforce access control through those sandboxes.

Adopt the software first and let the hardware follow your next server purchase

OpenShell is available on GitHub and NVIDIA's developer pages, and because it is open source, other chipmakers can extend it. NVIDIA names Arm and Intel. Sentry depends on NVIDIA's DOCA software and BlueField-4 hardware. NVIDIA says organizations can deploy pieces of the platform to fit their needs, so a team can put OpenShell in front of one agent this quarter and plan the hardware watchdog for its next server purchase.

Salesforce and NVIDIA connected OpenShell to Slack, so a team can review an agent's activity and approve or reject its requests for extra permissions inside Slack.

Competitors at one table are the strongest signal in the announcement

CrowdStrike and Palo Alto Networks compete for the same security budget, and both appear in NVIDIA's launch summary. ServiceNow, which competes with Salesforce for the same workflow budget, appears as well. SAP is building OpenShell into its Joule Studio and contributing engineering work to the project. Citi and JPMorganChase are working with NVIDIA on shared open source agent safety tools, and energy companies including NextEra Energy and Hitachi Energy are working with the platform.

NVIDIA started theOpen Secure AI Alliance with more than 120 organizations (NVIDIA, 2026), and the Linux Foundation governs it. The alliance asks policymakers to treat open models and security tools as defensive assets, and NVIDIA's founding blog post presents the launch as a response to regulators working out how to handle AI safety. I hope regulators and auditors read a shared set of controls that anyone can inspect as evidence they can test, because any rule about agent behavior needs a place where that behavior is recorded and stopped.

"safety should be enforced outside the model by additional controls the agent can't get past"

Mike Nicolls, president of SpaceXAI, said this in NVIDIA's launch release

The next invitations belong to the vendors behind your agents

The launch summary names 18 companies (NVIDIA, 2026) and the body of the release names dozens more. These vendors matter to many CIOs and CTOs, and none of them appears in the release:

• Google and Google Cloud are missing from the release and from the alliance's founding partner list.
• Amazon is on the founding partner list, and the release names Microsoft and Oracle Cloud Infrastructure as cloud partners without naming Amazon Web Services.
• OpenAI is missing from both the release and the founding partner list.
• Meta is missing from both as well.
• AMD goes unnamed, although the release says chipmakers such as Arm and Intel can extend OpenShell.
• The release says the platform brings together public-sector organizations, and it names no government agency among its partners.

A press release names partners on one date, and NVIDIA counts more than 120 organizations in the alliance (NVIDIA, 2026), so ask each vendor for its status in writing. The list has moved since July: Anthropic is absent from the alliance's founding partner list and named in the September release with an integration and an executive quote.

Two open questions belong on the proof of concept plan

• Ask for measured stopping times on your own workload, because NVIDIA states milliseconds (NVIDIA, 2026) and your agent traffic, rules, and network path set the number you will see.
• Decide who owns each agent's job description. In March I wrote that OpenShell enforces limits and cannot fix an agent whose task was written wrong, because a limit stops an action and a person writes the task.