Nvidia puts an enforceable boundary around autonomous AI agents
Nvidia’s Open Agent Safety Platform gives developers an open-source runtime that limits what long-running AI agents can access, adding a control layer outside the model itself. OpenShell turns operator instructions into enforced boundaries for software running on CPUs, while Sentry uses an out-of-band watchdog on BlueField-4 DPUs to monitor behavior and quarantine agents that cross those boundaries in milliseconds. The software can also extend to Arm and Intel platforms.