Nvidia has launched the Open Agent Safety Platform, a set of software and hardware safeguards designed to prevent autonomous AI agents from acting beyond their established boundaries.
The platform includes two controls. OpenShell provides what Nvidia calls a software-based “secure runtime boundary.” Sentry is a hardware-based “watchdog” that runs on the company’s BlueField-4 DPUs. Nvidia says Sentry can quarantine and stop an AI agent within milliseconds.
The company said recent security incidents showed that organizations need open, customizable tools. It argues that safety should not rely only on protections built into AI models and applications; software and hardware controls can add further barriers. Those controls complement model-level mechanisms that instruct AI systems not to behave in certain ways.
More than 100 customers, including SpaceXAI, Anthropic and Microsoft, are already using Nvidia’s software and hardware safeguards. Nvidia said the platform is intended to help organizations, researchers and public-sector groups share practices, align evaluation methods and cooperate internationally.
The announcement follows reports of AI agents operating beyond their intended controls. Nvidia’s approach draws on established security principles such as least privilege, isolation and detailed monitoring. CEO Jensen Huang said the company should accelerate work on AI safety as capabilities advance.
Comments
0No comments yet. Be the first to comment.