NVIDIA turns open AI security into a coalition with Microsoft and Cloudflare

The open-model fight is moving from weights to the full defensive stack around AI agents. On July 27, 2026, NVIDIA launched the Open Secure AI Alliance, a coalition meant to build and share open technologies for AI safety and cybersecurity. Initial partners include Microsoft, Cloudflare, CrowdStrike, Hugging Face, IBM, LangChain, the Linux Foundation, Palantir, Palo Alto Networks, Red Hat, Salesforce, ServiceNow, Siemens, Snowflake, and Thinking Machines Lab.

The alliance is framed as a practical answer to a specific security problem: defenders often need to inspect, adapt, and run AI tools on their own infrastructure. NVIDIA argues that an AI agent is not just a language model. It is a system made of identity, permissions, harnesses, guardrails, logs, and evaluation. In that framing, the safety question is not simply whether weights are open or closed, but whether the surrounding controls can be tested and improved by the people responsible for security.

NVIDIA points to the recent Hugging Face security incident as the example. According to the blog post, closed AI tools blocked essential forensic analysis because they could not distinguish attackers from defenders. Hugging Face then ran the open-weight GLM 5.2 model on its own infrastructure to analyze more than 17,000 actions and contain the intrusion. That number is the policy argument in miniature: when response speed matters, defenders may not be able to wait for permission from a closed provider.

The initial technical contributions map to the agent stack. NVIDIA is contributing open models, model weights, data, and agent harness research, including the NVIDIA Labs Object-Oriented Agent project for testing, tracing, auditing, and governing agent behavior. HPE is tied to SPIFFE/SPIRE work for zero-trust identity. Hugging Face offers Safetensors as a model-weight format designed to avoid remote code execution. IBM and Red Hat point to signed patch workflows, while Microsoft contributes MDASH, a multi-model agentic scanning harness.

The policy stakes are immediate. U.S. lawmakers are debating whether frontier open models should face stronger restrictions, while industry players are trying to define openness as part of national cyber defense. The alliance will be judged less by its membership list than by whether it ships usable shared datasets, attack simulators, red-team tools, and deployment patterns that security teams actually adopt.