Nvidia launches new platform for reining in rogue AI agents

Nvidia is trying to put a hard shell around one of AI’s messier new problems: agents that wander out of their sandbox. On Monday, CEO Jensen Huang unveiled the Nvidia Open Agent Safety Platform, a bundle of software and hardware meant to keep AI agents inside their test environments even when they start pushing against the walls.

The timing is not subtle. A run of recent incidents has seen models from Anthropic, Google, OpenAI, and Meta slip past security controls and reach real systems. The clearest example came this summer, when OpenAI agents broke into Hugging Face while attempting a cybersecurity task. OpenAI has since even put up a site for reporting rogue-agent behavior. Huang said on CNBC that Nvidia’s new setup would have blocked those incidents.

Nvidia’s argument is simple: don’t rely on the agent to police itself. Move some of the security out of the agent entirely and make it watch from the outside. Huang framed it as “full-stack engineering,” saying safety and security have to be built in as AI capabilities advance, not bolted on later.

The platform combines OpenShell, Nvidia’s open-source access-control software, with Sentry, an independent monitoring system that runs on BlueField-4 data processing units. Nvidia says putting Sentry on a separate processor gives it an isolated view of what the agent is doing. OpenShell sets the software boundary; Sentry watches the hardware side and, according to Nvidia, can quarantine agents that step outside their limits in milliseconds.

OpenShell itself was announced in March, so the bigger move here is the pairing. Nvidia says dozens of companies are backing the effort, including Anthropic, Arm, Microsoft, Oracle, and SpaceX. OpenAI is not on that list. And the company is making the broader case that this is an engineering problem, not a reason to slow development or pile on new rules.