Nvidia stakes a claim on AI safety infrastructure with 100+ partners on board

In analysis terms, Nvidia's new Open Agent Safety Platform is as much a business strategy as a safety product: give away the software that contains AI agents, then sell the hardware that watches them, tying agent safety directly back to Nvidia's own chip sales rather than leaving it to a third party.

The free half, OpenShell, is an open-source agent runtime released under the Apache 2.0 licence that runs each agent in its own sandbox and checks that its permissions cannot combine into something the operator never intended, according to Nvidia. It is tuned for Nvidia's own Vera CPUs, though it also runs on Arm and Intel hardware.

The proprietary half, Sentry, is where the commercial logic sharpens: it runs on Nvidia's BlueField-4 data-processing unit, kept deliberately separate from the host machine so a compromised agent cannot switch off its own monitoring, and Nvidia says it can quarantine a rogue agent within milliseconds.

Sentry is not open source, though it exposes open application programming interfaces, and is offered as an optional add-on — meaning every enterprise customer who wants the fastest containment layer is nudged toward buying Nvidia's own silicon to run it on.

More than 100 organisations are already involved, Nvidia said, including Anthropic, Microsoft, SAP, Scale AI and JPMorgan Chase — a list that spans AI labs, software vendors and, notably, a major bank whose presence signals real enterprise appetite for this kind of oversight rather than just developer curiosity.

Nvidia's vice president of enterprise AI, Justin Boitano, said the system could have prevented the breach of Hugging Face's infrastructure carried out earlier this year by another company's agents, per TechCrunch — a pointed comparison that puts a concrete incident behind an otherwise abstract safety pitch.

The capability claims, though, come entirely from Nvidia and its partners and have not been independently tested, which matters for a company whose core business is selling the chips its own safety product happens to run fastest on. No independent auditor has yet confirmed Sentry's claimed millisecond quarantine time under real conditions.

The partner integrations add weight to the pitch: Salesforce has wired OpenShell into Slack so users can approve or reject an agent's requests for extra permissions, and Anthropic has paired the platform with its Claude Managed Agents service.

For the banks and large enterprises among the 100-plus partners, the pitch is straightforward even if unverified: a hardware-isolated safety net that a compromised agent cannot simply talk its way around, sold by the same company that already supplies the chips running their AI workloads.

Whether rivals such as AMD or cloud providers offer a comparable independent safety layer of their own will likely determine how much leverage this gives Nvidia beyond the chip market it already dominates.