NVIDIA is responding to the industry risks posed by out-of-control autonomous AI agents with a new security architecture.

On Monday, NVIDIA (NASDAQ: NVDA) released a dual-layer AI safety platform called the "Open Agent Safety Platform," which includes two open-source software tools, OpenShell and Nvidia Sentry. These tools can run on NVIDIA hardware to control AI agents' access permissions in real time and shut them down when they violate rules.

NVIDIA said the system can isolate abnormal agents within milliseconds, and claimed that if relevant labs had already deployed this technology, the July incident in which an OpenAI model attacked Hugging Face could have been prevented.

This move is also NVIDIA's latest step to expand beyond its chip business into the AI software security field. At the same time, CEO Jensen Huang has recently repeatedly described AI safety as an engineering problem, stressing that AI systems must undergo rigorous safety testing.

Dual-layer architecture monitors in real time and isolates abnormal agents within milliseconds

Justin Boitano, NVIDIA's vice president of enterprise AI, said at a pre-launch media briefing that the Open Agent Safety Platform can monitor the behavioral boundaries of AI agents in real time and intervene immediately once violation rules are triggered. "Based on what we know, this new safety platform could have stopped that intrusion," he said.

The platform consists of two open-source tools: OpenShell is responsible for controlling the resources and execution permissions that AI agents can access, while Nvidia Sentry monitors agent behavior and identifies potential threats. Both can run on NVIDIA hardware. NVIDIA said the architecture is designed to help enterprises and AI labs test advanced AI systems within stricter safety boundaries while avoiding slowing down research and development because of security risks.

The background to NVIDIA's release is the continued industry attention on security incidents involving autonomous AI agents. In July, a security incident involving an OpenAI model at Hugging Face further intensified concerns about autonomous AI agents going out of control and sparked discussion about whether AI development should be slowed.

Boitano declined to comment on whether OpenAI and Anthropic plan to use the system to monitor their training processes, saying such decisions depend on the companies themselves.

Jensen Huang: AI safety is first and foremost an engineering problem

On regulatory issues, Huang has recently emphasized repeatedly that AI safety should be addressed more through technical means rather than relying solely on new regulatory frameworks. He has also publicly questioned warnings from some AI developers that AI could threaten human survival, while stressing that AI systems must still undergo rigorous safety testing.

Judging from this product rollout, NVIDIA is trying to further embed AI safety capabilities into the infrastructure for model operation and agent execution, reducing the safety risks of autonomous agents through permission controls and anomaly isolation. This also means NVIDIA's business reach is extending further from AI chips into software and security infrastructure.