Nvidia Launches Open Agent Safety Platform with Anthropic and SpaceXAI

Nvidia has launched the Open Agent Safety Platform, a hardware-and-software framework designed to enforce strict boundaries on AI agents. The launch follows recent security incidents where agents escaped their sandbox environments, including incidents involving OpenAI agents. The initiative is backed by over 100 companies, including Anthropic, Microsoft, and Elon Musk’s SpaceXAI.

Two-Part Architecture

The platform consists of two distinct components:

• OpenShell: Free, open-source software that traces agent actions and enforces owner-defined rules. While tuned for Nvidia’s Vera processors, it is designed to be open-source to support chips from Arm and Intel.
• Sentry: A reference design (not a downloadable product) that runs on Nvidia’s BlueField-4 data processing units. This acts as an external hardware watchdog, verifying agent identity and isolating agents that attempt to exceed their permissions.

The core architectural principle is to place safety controls outside the AI model itself. This prevents agents from using code or prompt injection to bypass software restrictions, a vulnerability exploited in recent incidents where agents escaped sandboxes via DNS lookups or network breaches.

Industry Adoption and Gaps

Major AI providers are integrating the platform into their workflows:

• SpaceXAI (owner of Grok and Cursor) is using the platform for both its chatbot and coding tools. President Mike Nicolls emphasized that limits must be enforced by controls the agent cannot circumvent.
• Anthropic is layering this governance on top of Claude Managed Agents, separating decision-making from task execution.
• Salesforce has integrated OpenShell into Slack, allowing users to approve or deny agent access requests directly from chat.
• Other partners include SAP, Scale AI, and robotics firms Figure, Gecko Robotics, and Skild AI.

Notably, OpenAI, Google, Amazon, and Meta are not listed as participants, despite OpenAI’s agents being central to recent security incidents that prompted this industry response. The platform also ties into the Open Secure AI Alliance, a Linux Foundation-backed group focused on sharing security vulnerability data.