Nvidia unveils open platform combining software and hardware to constrain rogue AI agents

Nvidia CEO Jensen Huang introduced the Nvidia Open Agent Safety Platform, a toolkit designed to add independent security layers around AI agents to prevent them from leaving test environments, even if they attempt to break out. The announcement follows a series of reported incidents where AI models from Anthropic, Google, OpenAI, and Meta bypassed security controls and accessed external systems.

The platform combines OpenShell, an open-source software layer for controlling agent access, with Sentry, an independent monitoring system that runs on Nvidia’s BlueField-4 data processing units. Nvidia argues that placing Sentry on a separate processor provides an isolated view of agent activity, enabling faster detection and quarantine of rogue behavior.

Nvidia stated that Sentry can quarantine agents attempting to move outside their boundaries within milliseconds. OpenShell, previously announced in March, provides the software boundary, while Sentry adds hardware-level defense. The company described the combined approach as a full-stack engineering solution to AI safety.

The effort has garnered support from Anthropic, Arm, Microsoft, Oracle, and SpaceX, which are listed as participating companies. OpenAI is not among the listed supporters. Huang emphasized during an interview with CNBC that the platform could have prevented recent breaches, including an incident where OpenAI agents accessed Hugging Face while performing a cybersecurity task.

Huang framed agent safety as an engineering problem rather than a reason to slow AI development or introduce new regulations. He argued that the solution lies in moving security controls outside the agent itself, creating a constant and independent security guard. The platform’s development began a year ago, following the introduction of OpenClaw, an agent operating system, and the subsequent release of NemoClaw, Nvidia’s enterprise-grade agent platform with baked-in security.