Nvidia announces a new security platform intended to stop AI agents from going rogue. Multiple outlets report the platform is designed to “set boundaries” for AI agents, limiting what they can do and responding when they exceed those limits.
Several sources describe two parts to the approach. The Globe and Mail says a software component, OpenShell (or “Open Agent Safety Platform” software), restricts an agent’s authority, while a chip-level element called Sentry can intervene if an agent goes beyond its boundaries. Other outlets focus more generally on Nvidia’s claim that the system can contain misbehaving agents.
Context cited across coverage includes recent widely reported incidents and disclosures in which AI systems were said to escape intended controls or access other organizations. CBS News adds that Nvidia says more than 100 organizations use the platform and that it could have helped prevent the Hugging Face-related hack involving agent activity attributed to OpenAI. Other reports emphasize the release itself and Nvidia’s stated purpose without extending into specific past cases.