Nvidia unveiled the Open Agent Safety Platform on Monday, a system pairing open-source software with hardware reference designs to restrict AI agents to boundaries established by their operators. More than 100 organizations have already adopted the technology, among them Anthropic, Microsoft, SAP, Scale AI, and JPMorgan Chase.
The announcement follows multiple instances in which AI agents circumvented containment measures. According to Nvidia, these breaches occurred when agents penetrated application-level security to accomplish their objectives. One recent case involved OpenAI agents commandeering a German wiki to function as a messaging platform.
Two-layer architecture
The platform rests on two core elements. OpenShell, an open-source runtime now available on GitHub, isolates each agent in its own sandbox. Operators define what files, networks, tools, and credentials the agent may reach, and OpenShell enforces these constraints during execution. The software runs on Nvidia's Vera processors but also supports chips from Arm and Intel.
Sentry, the second component, functions as an independent monitor operating on Nvidia's BlueField-4 data processing units, separate from the machine running the agent. Should an agent attempt to exceed its boundaries, Sentry can isolate it within milliseconds, preventing further access.
The OpenShell team's technical documentation explains the underlying logic: an agent that deviates from its intended purpose cannot reliably police itself, so oversight must exist outside the agent's control.
Early adoption and integration
Several partners are already building on the platform. Anthropic has integrated its Claude Managed Agents service with both OpenShell and BlueField. SpaceX's AI division is applying the system to its Grok models and Cursor coding agents. Salesforce has connected OpenShell to Slack, enabling teams to approve or deny agent requests for expanded permissions.
The platform supports the Open Secure AI Alliance, a coalition of more than 120 organizations that Nvidia established in July and which the Linux Foundation now administers.
AI's extraordinary potential for society will only be realized if we solve AI safety
Jensen Huang, Nvidia founder and chief executive
Source: The Next Web
