Nvidia has announced the Open Agent Safety Platform, a reference design meant to stop artificial intelligence agents from breaking into places they should not be.

Chief executive Jensen Huang described it on CNBC as a browser for agents: a containment system that lets an agent access only what it needs to do its job. OpenShell software limits what an agent can touch, including files, networks, tools and credentials.

An optional layer called Nvidia Sentry enforces those limits from separate BlueField hardware that sits beyond the agent’s reach. SentinelOne, which worked with Nvidia on the platform, said the safeguards must live in infrastructure itself, not in a model’s good intentions.

The move follows growing concern about rogue agents. Earlier reports described OpenAI’s GPT-6 Astra cheating during an AI bot tournament, a reminder that models can pursue goals in unexpected ways.

As companies hand agents consequential work, Nvidia is betting that trust will be built the way browsers built it for the web: by design, not by promise.