Nvidia Adds Layered Security for AI Agents

Nvidia introduced the Open Agent Safety Platform to contain autonomous AI agents after several recent cases in which models bypassed sandbox restrictions. OpenShell runs agent workloads inside policy-controlled environments and limits access to files, credentials, processes and network services. A separate supervisor can inspect outbound requests and enforce fine-grained permissions even when agents launch generated code or create sub-agents. Nvidia also added Sentry, a hardware-level monitor running on BlueField-4 DPUs that can independently watch agent behaviour and intervene when activity moves outside expected boundaries. Nvidia says the architecture could have prevented the Hugging Face breach involving OpenAI agents. The company is presenting the system as a layered security model that keeps critical controls outside the AI agent itself.

Digital Trends, Sudhanshu Kumar Mangalam – Nvidia has a fix for rogue AI agents to prevent incidents like the Hugging Face hack – 28th September 2026