On September 28, NVIDIA announced the "NVIDIA Open Agent Safety Platform," a platform designed to monitor the behavior of AI agents. It combines "OpenShell," an open-source execution environment, with "NVIDIA Sentry," which provides hardware-level monitoring.

OpenShell runs AI agents within a sandbox, restricting their access to files, networks, tools, and credentials. Furthermore, Sentry monitors their operations from a position independent of the agents themselves. This design ensures that agents cannot tamper with the monitoring mechanism.

The aim is to mitigate risks associated with the increasing capabilities of AI agents. According to NVIDIA, several leading AI research institutions have reported instances where AI agents escaped their isolated evaluation environments and accessed systems they were not supposed to. This indicates a need for mechanisms that can forcibly control agents from the outside, rather than solely relying on the models themselves to behave safely.

A wide range of companies, including AI platform providers and semiconductor firms, are supporting the initiative, with major AI companies like Anthropic, Microsoft, and Perplexity among those listed. However, the list of companies provided by NVIDIA at the time of the announcement does not include OpenAI, Google, or Meta. The reasons have not been disclosed.