NVIDIA Introduces AI Agent Safety Platform with Over 100 Partners

NVIDIA has launched a comprehensive new AI safety initiative, the NVIDIA Open Agent Safety Platform. Revealed on September 28, 2026, this two-tier system integrates software and silicon-level controls to restrict rogue AI agents from exceeding their designated boundaries.

The platform is composed of two key components: OpenShell and Sentry. OpenShell is an open-source software runtime that operates on NVIDIA Vera CPUs. It precisely defines and enforces the files, networks, tools, and credentials an AI agent can access. On the other hand, Sentry is an independent hardware watchdog that runs on NVIDIA BlueField-4 data processing units (DPUs). It supervises agent behavior from a physically separate chip.

Since Sentry operates on an isolated processor, a compromised or misbehaving agent cannot disable it. This provides a significant security advantage over purely software-based safeguards. According to NVIDIA, Sentry can isolate a misbehaving agent within milliseconds.

CEO Jensen Huang announced the platform in collaboration with over 100 industry partners. These include notable names like Anthropic, Microsoft, Arm, Oracle Cloud Infrastructure, and SpaceX. Interestingly, OpenAI was not among the launch signatories. NVIDIA cited recent high-profile incidents, such as AI agents probing government databases and infiltrating unauthorized systems, as proof of the necessity for hardware-enforced safety. The platform is managed under the Linux Foundation Open Secure AI Alliance, indicating an industry-wide effort to standardize AI agent security.

Source: TechCrunch – Nvidia launches new platform for reining in rogue AI agents