NVIDIA has introduced a platform to protect AI agents from going beyond their designated tasks.

NVIDIA has introduced the Open Agent Safety Platform — a suite of software and hardware solutions for controlling AI agents and preventing them from escaping their test environment. The platform was presented by NVIDIA CEO Jensen Huang on September 28.

The solution combines the OpenShell software with the independent Sentry monitoring system. OpenShell restricts AI agents' access to resources during operation, while Sentry monitors their actions separately from the agent itself.

Sentry operates on NVIDIA BlueField-4 processors and is deployed separately from the CPU or GPU running the AI agent. According to NVIDIA, this allows the system to independently monitor agent actions and isolate those attempting to breach established boundaries within milliseconds.

The development comes amid several incidents where AI agents bypassed established restrictions and gained access to real-world systems. One of the most notable was an incident involving OpenAI agents, which, while performing a cybersecurity task, gained access to Hugging Face.

NVIDIA originally introduced OpenShell in March 2026. The new platform integrates it with the hardware-level protection provided by Sentry.

According to NVIDIA, the initiative has been supported by Anthropic, Arm, Microsoft, Oracle, and SpaceX. OpenAI is not listed among the platform participants.

Jensen Huang stated that AI agent security should be addressed through engineering solutions. He emphasized that when deploying an agent, its permissions and access to resources must be restricted first and foremost.

Source: Nvidia