Nvidia launches Open Agent Safety Platform to lock down rogue AI agents

Nvidia has launched an Open Agent Safety Platform to prevent rogue AI agents from breaking out of their test environments and gaining access to real systems. The platform combines OpenShell 0.1.0, an agent runtime announced by Nvidia at GTC in March, with Nvidia Sentry, a watchdog service running on the company's BlueField-4 data processing units (DPUs).

OpenShell adds a policy prover that checks agent permissions to ensure they don't combine into unintended actions, such as hacking. Nvidia's vice president of enterprise AI, Justin Boitano, stated that model-level safeguards alone cannot govern agents' access or actions. The new platform aims to address this fundamental hurdle by enforcing policy outside the agent.

OpenShell operates each agent in a kernel-isolated sandbox with no network access except through a supervisor outside the workload. The policy prover, running roughly twice as fast as traditional methods, ensures permissions stay within the operator's intended boundaries. Nvidia Sentry provides an additional hardware layer, running on BlueField-4 in a trust domain separate from the host, to quarantine agents in milliseconds.

This combination of OpenShell and Sentry allows Nvidia to mediate and enforce how agents behave, addressing the limitations of model-level safety approaches.