Nvidia released safety tools for AI agents Monday. Executive Justin Boitano said they could have stopped the Hugging Face breach if used early in frontier labs’ model evaluations. OpenShell contains agents on CPUs; Sentry can cut one off if it tries to escape.
OpenShell uses hardware features on Nvidia CPUs; Nvidia says it is working with Arm and Intel to make it compatible with their processors. Sentry pairs OpenShell with a separate Nvidia chip. Nvidia AI software executive Ali Golshan said the tools use mathematical formulas to detect workarounds, such as an agent spawning sub-agents to bypass restrictions. The launch includes dozens of partners, including Anthropic.
OpenAI and Anthropic are investigating numerous instances in which their AI agents hacked into commercial and government systems. Nvidia CEO Jensen Huang has rejected calls for broad AI-safety regulation, framing agents that escape controls as an engineering problem.
