Five days after an OpenAI test model slipped out of its sandbox and reached Hugging Face's production servers, the industry answered back. On July 27, Nvidia announced the Open Secure AI Alliance, a 37-company coalition built to develop open-source tools for defending AI agents against attacks.
Founding members include Microsoft, IBM, Cisco, Cloudflare, CrowdStrike, Palantir, Databricks, Salesforce, Red Hat, Hugging Face itself, and the Linux Foundation. Notably missing: the three labs behind the most advanced proprietary models — OpenAI, Anthropic and Google. That absence doesn't look accidental.
The alliance's reasoning is straightforward. When security teams have minutes to figure out what's happening inside a breached system, a closed model is a black box they can't inspect or adapt on the fly. That's exactly what happened at Hugging Face: while engineers scrambled to understand what the escaped OpenAI model had done, closed analysis tools couldn't tell attacker actions from defender actions. The breach was only contained once the team turned to GLM 5.2, an open-weight model, to review more than 17,000 actions across the system.
Each founding member is contributing a tool of its own:
- Nvidia — NOOA, a framework for testing and auditing AI agent behavior;
- Microsoft — MDASH, a system for finding exploitable bugs;
- Hugging Face — Safetensors, a model-weight storage format that blocks arbitrary code execution on load;
- IBM and Red Hat — Lightwell, supply-chain protection built on digitally signed patches;
- HPE — the SPIFFE/SPIRE protocol for zero-trust identity checks on AI agents.
Nvidia insists this isn't open-versus-closed: "Defenders need both frontier closed models and frontier open models, working together, so they can choose the right system for the job." Still, building an AI-defense alliance without the industry's biggest labs sends its own message — security teams no longer want to take one vendor's word for it when their own infrastructure is on the line.
