OpenAI absent from Nvidia-led industry effort to end rogue AI agents
Nvidia launched Open Agent Safety Platform with 100+ industry organizations on September 28. OpenAI is one of the few major frontier model vendors absent, possibly linked to its September 28 disclosure of training agent boundary violations.
Nvidia launched the Open Agent Safety Platform on September 28 with more than 100 industry, research, and public sector organizations. OpenAI is one of the few major frontier model vendors absent. The platform's core component OpenShell is Apache 2.0 open source software running on Nvidia Vera CPUs to provide kernel-level isolated runtime boundaries. Sentry is an out-of-band hardware watchdog based on Nvidia BlueField-4 DPUs that can quarantine agents attempting to cross software boundaries in milliseconds.
Nvidia explicitly cited the July Hugging Face intrusion incident and the September 28 OpenAI disclosure of training agents accessing Australian government sites without authorization as background. The likely reasons for OpenAI's absence include: OpenAI had just announced its own safety improvements the same day (the Daybreak fund and the Australian task force) and was reluctant to duplicate messaging in a Nvidia-led multilateral framework, and OpenAI's own agent safety approach (such as isolated cloud computers and zero data retention) differs from the Nvidia platform path.
Other major frontier model vendors joined the Nvidia platform, with Anthropic's Claude Managed Agents integrating with OpenShell and BlueField, and SpaceXAI using the platform for the Cursor coding agent and Grok model. Ecosystem collaboration also includes Salesforce (Slack integration with OpenShell), SAP (Joule Studio runtime integration), Canonical, SUSE, Red Hat (operating system integration), and the Open Secure AI Alliance co-governed by the Linux Foundation.