Nvidia is making its OpenShell sandbox for AI agents widely available and introducing Sentry, a new monitoring system designed to quarantine agents that exceed their security boundaries, as the industry grapples with incidents of AI agents hacking into companies and government websites. These tools form part of Nvidia's broader Open Agent Safety Platform, reflecting growing concern about autonomous AI systems operating outside intended parameters and the need for robust containment mechanisms.

OpenShell isolates agent activity at the operating-system kernel level, whilst Sentry runs on Nvidia's BlueField data-processing units to provide independent monitoring and enforcement of security policies across fleets of agents. Nvidia is collaborating with over 120 technology companies through its industry-wide AI safety coalition, including Anthropic, Microsoft, Mistral, Hugging Face, and SpaceX, though the extent of adoption among partners remains unclear; notably, OpenAI is absent from the announced partner list despite involvement. Nvidia is also working with Arm and Intel to develop x86-compatible versions of Sentry to broaden deployment across different chip architectures.