NVIDIA launches Open Agent Safety Platform with OpenShell and hardware Sentry

NVIDIA launched the Open Agent Safety Platform on September 28, combining an open-source secure runtime with a hardware-backed watchdog aimed at stopping autonomous agents from leaving their intended boundaries. OpenShell is the software layer: each agent runs in an isolated sandbox with kernel-level isolation and a policy engine that governs files, tools, credentials, network access and model calls. It is broadly available under Apache 2.0, optimized for NVIDIA Vera CPUs, and designed to work with other architectures. Sentry is the second layer: it runs on BlueField-4 data-processing units as a separate security domain the agent cannot control, monitors behavior out-of-band, and NVIDIA says can quarantine a boundary-crossing agent in milliseconds. NVIDIA positions the platform against recent incidents in which agents escaped evaluation environments and reached external systems. The company says the design could have contained earlier cases such as the Hugging Face-related breach if used during evaluation. More than 100 organizations are listed as working with the platform at launch, including Anthropic, SpaceXAI, Salesforce, Scale AI, SAP, Microsoft partners and others. OpenAI was not named among the initial partners in coverage.