Nvidia opens OpenShell to all, adds Sentry for AI agents

Nvidia's move comes after months of disclosures from frontier AI labs about agents hacking into other companies and probing official US and Australian government websites. Nvidia already had an agent security sandbox, OpenShell, announced at its GTC Conference in March, and in July it launched an industry-wide AI safety coalition that now comprises more than 120 companies, with a program called the Shared AI Findings Exchange (SAFE) intended to be governed independently. The new Open Agent Safety Platform wraps OpenShell and Sentry together, and Nvidia says it is working with Arm and Intel to make Sentry run on the x86 chip architecture, which would let it run on any architecture.

Nvidia's list of collaborators spans chip, cloud, security and enterprise names such as Anthropic, Cisco, CoreWeave, CrowdStrike, Dell Technologies, Hugging Face, JPMorganChase, Mistral, Microsoft and Palantir. Nvidia says SpaceXAI is using the Open Agent Safety Platform for its Cursor agents and Grok models, and that Anthropic and Nvidia are building security into Claude Managed Agents, while Salesforce, Scale AI and SAP are integrating OpenShell to some degree. OpenAI's absence from the announcement, despite both companies saying it is part of the effort, leaves the scope of adoption unclear.

Nvidia frames the tools as a way to give security teams collective policy over fleets of agents. Justin Boitano, Nvidia's vice president and general manager of enterprise computing, says agents are very creative at finding ways to achieve their goals, and that with this, agents only have access to the intent the security team wants them to have. The test is likely to be whether the named partners actually deploy OpenShell and Sentry rather than simply appear on a list, and how quickly Sentry reaches x86 chips through the Arm and Intel work.