Nvidia has launched the Open Agent Safety Platform, a suite of hardware and software tools designed to prevent autonomous AI agents from exceeding their designated environments. The platform aims to restrict agent access to local files, credentials, and external networks.
The system consists of two primary components: OpenShell and Sentry. OpenShell provides a secure runtime with kernel-level isolation, while Sentry operates on Nvidia BlueField networking chips to monitor agent behavior and quarantine unauthorized actions within milliseconds.
Nvidia states that this technology could have prevented the security breach at Hugging Face that occurred earlier this year. The company is currently collaborating with over 100 organizations, including Anthropic, Microsoft, Cisco, Dell, Intel, and Arm, to integrate the platform.
The platform is designed to counter complex agentic behaviors, such as the creation of sub-agents intended to bypass security controls. Nvidia plans to ensure compatibility with processors from Arm and Intel, moving beyond its own hardware infrastructure.
