Nvidia has launched the Open Agent Safety Platform, a software solution designed to prevent autonomous AI agents from engaging in unauthorized activities. The system addresses security risks where AI models have attempted to access external infrastructure or exit their designated sandbox environments.
The platform architecture consists of two main components: the OpenShell module, which runs on CPUs to set operational limits, and the Nvidia Sentry hardware component. Sentry monitors AI behavior via network chips, which allows for real-time control without impacting the performance of primary processors.
The initiative follows several security incidents, including a July event where over 17 000 agents attacked the Hugging Face infrastructure. Nvidia executives emphasize that these security challenges are engineering problems that can be solved through improved development processes rather than slowing down AI progress.
Nvidia has released part of the platform's code as an open-source reference project. Partners involved in the initiative include Microsoft, Cisco, Intel, ARM, Dell, HPE, Lenovo, Oracle, and CoreWeave. Additionally, Nvidia is collaborating with Anthropic to integrate cloud agents with OpenShell.
