Nvidia tells that an isolated unit must monitor AI agents

Gilad Shainer is a networking leader at Nvidia. Shainer told theCUBE that AI agents (software that does tasks) must have a monitor that they cannot touch. Shainer tells that Scale-In, which Nvidia introduced in August, does this. It uses BlueField-4, a data processing unit (hardware for network, storage and security work). Nvidia gave no test results.

How it works

Scale-In operates on BlueField-4 and DOCA, a software kit for that unit. The unit works out of band, that is, it is not in the path of the agents. OpenShell, an open-source tool, controls the tasks and data that agents can use. Sentry operates on BlueField-4 and can isolate an agent that breaks its limits.

The numbers

The source gives no test results, speed numbers or prices. Nvidia is the source of all claims. The steps are:

• August: Scale-In
• September: Open Agent Safety Platform, with OpenShell and Sentry
• October: OpenShell Reference Architecture

What is not known

The source does not tell how correctly this operates in use or the price. It does not tell if the monitor operates on hardware from other makers.

Insight: Nvidia sells the agent monitor with its own hardware

Each step of Nvidia agent security, from August to October, operates on BlueField-4. A stack that Nvidia showed on 9 November also has OpenShell. Its test numbers come from Nvidia too. This reads as a toll gate: the maker that sells the hardware can also set the rules for safe agents. Customers must find out if they can use a monitor from a different vendor.