Nvidia has introduced the Open Agent Safety Platform, a new toolkit designed to prevent AI agents from escaping their designated testing environments. The system aims to address recent security breaches involving models from companies like OpenAI, Anthropic, Google, and Meta.
The platform integrates OpenShell, an open-source software for access control, with Sentry, a monitoring system that operates on Nvidia BlueField-4 data processing units. By isolating security monitoring on a separate processor, the platform can detect and quarantine agents attempting to breach boundaries within milliseconds.
Development of this security initiative began a year ago, following the release of the OpenClaw operating system. In March, Nvidia previously launched NemoClaw, an enterprise-grade platform that incorporated initial security features.
Several major technology companies have committed to supporting the platform, including Anthropic, Arm, Microsoft, Oracle, and SpaceX. OpenAI is currently not among the participating organizations.
Nvidia CEO Jensen Huang stated that the company does not support slowing AI development or imposing new industry regulations. Instead, the firm advocates for full-stack engineering solutions to manage AI safety independently of the agents themselves.
