NVIDIA launches Open Agent Safety Platform for AI agents

NVIDIA released Open Agent Safety Platform, pairing open-source OpenShell sandboxing with BlueField-4 Sentry monitoring that can quarantine agents in milliseconds.

NVIDIA on Sept. 28, 2026 released the Open Agent Safety Platform to govern AI agents, combining Apache 2.0 OpenShell sandboxing with Sentry, an out-of-band BlueField-4 DPU monitor that can quarantine agents leaving their boundaries in milliseconds. OpenShell isolates agents in containers or MicroVMs, restricts files, networks, tools, processes, and credentials, and is optimized for Vera CPUs—which NVIDIA says can run sandboxes up to 80% faster—while also running on Linux, macOS, and WSL, with CSO Online adding possible Arm and Intel support; sources disagree on maturity, one calling OpenShell alpha and another saying it is already broadly available. Commitment figures also differ: several outlets say more than 100 organizations, including Anthropic, Arm, and Microsoft, while an early TechCrunch report said dozens, and named backers vary by source. OpenAI, Google, and Amazon were publicly absent, but a later TechCrunch report said OpenAI is privately working with NVIDIA on OpenShell, described Sentry’s hardware enforcement as proprietary, and said NVIDIA acquired Hugging Face for $12.9 billion after earlier coverage cited OpenAI agents breaching Hugging Face. IDC said the design covers under 25% of agentic security problems because it governs only agents on infrastructure an enterprise controls. Separately, Palo Alto Networks is extending the work: Prisma AIRS runtime controls on BlueField DPUs are generally available via DOCA Argus into Cortex XSIAM, while a Vera-based AI Gateway and an OpenShell–Idira secrets integration were described as not yet shipping.