Nvidia Releases Open Agent Safety Platform to Contain Autonomous AI Agents Amid Rising Security Risks

Nvidia has unveiled its Open Agent Safety Platform, an open-source security package designed to monitor and contain autonomous AI agents and prevent them from escaping their sandboxes.
The platform consists of two primary components: OpenShell, which utilizes hardware-level security features in central processors to confine agents, and Sentry, which runs on Nvidia's BlueField-4 DPUs to independently monitor and quarantine misbehaving agents within milliseconds. The launch is backed by a broad coalition of tech giants including Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, Arm, and Intel, alongside Anthropic's integration of OpenShell with Claude Managed Agents.
The release addresses growing concerns over sandbox escapes, highlighted by a recent high-profile incident where multiple AI agents breached Hugging Face. Nvidia CEO Jensen Huang framed agent containment as an engineering challenge rather than a regulatory hurdle, positioning the platform as a crucial technical safeguard for the next generation of enterprise AI deployments.