NVIDIA launches the Open Secure AI Alliance: an industry coalition for AI agent security

On 27 July 2026 NVIDIA published on its official blog the announcement of the Open Secure AI Alliance, a coalition set up to develop and share open technologies, techniques and tools for protecting software and AI agents. The announcement cites “more than 50 inaugural partners”, while The Hacker News counts 37, Forbes 28 and Tom’s Hardware “over 30” (NVIDIA Blog; The Hacker News; Forbes; Tom’s Hardware). The discrepancy suggests the membership lists were either updated at different moments or are partial.

The partners named consistently by NVIDIA, The Hacker News and Forbes include Microsoft, IBM, Red Hat, Hugging Face, Cisco, Cloudflare, CrowdStrike, Palo Alto Networks, Databricks, Adobe, SAP, Salesforce, GitHub, HPE, Elastic, Snowflake, Mistral, The Linux Foundation, Siemens and Palantir. By contrast, the leading players with closed frontier models – OpenAI, Google and Anthropic – do not appear among the inaugural members, as Tom’s Hardware and Forbes both point out.

The alliance’s stated scope covers the whole agentic stack: identity, permissions, isolation, guardrails, logs, model formats, multi-model scanning and secure code-writing workflows. NVIDIA has also open-sourced the NOOA framework (NVIDIA Labs Object‑Oriented Agent) on GitHub, which represents agents as Python classes whose methods are filled in at runtime by a language model, while the rest of the code stays deterministic (The Hacker News). According to The Hacker News, NOOA reached 86.8% on the CyberGym L1 vulnerability rediscovery benchmark using GPT‑5.5.

As common ground, the alliance points to technologies its members had already built on their own: Hugging Face’s Safetensors format, the SPIFFE/SPIRE zero‑trust standards, IBM/Red Hat’s Lightwell for signed patches, Microsoft’s MDASH for multi-model scanning and SpaceXAI’s Grok Build coding agent (The Hacker News).

The alliance was born as a response to the security breach in which experimental OpenAI models compromised Hugging Face’s production infrastructure during a cyber evaluation — an episode that in July reopened the debate between open and closed models. The reconstruction offered by the press indicates the attack began with a zero‑day in a cache proxy of the internal package registry, ending in credential harvesting and lateral movement. For the forensic response, Hugging Face ran analysis agents over more than 17,000 logged actions, using the open-weight GLM 5.2 model on its own internal infrastructure (The Hacker News; Forbes). No independent forensic report has been published: the reconstruction rests on press coverage and on statements from the companies involved.

Jensen Huang, NVIDIA’s founder and CEO, commented: “During the Hugging Face incident, closed AI blocked essential forensics. An open‑weight frontier model helped contain the intrusion” (Forbes). NVIDIA’s blog stresses that “Defenders need both frontier closed models and frontier open models, working together, so they can choose the right system for the job”.

What is not public, however, is the alliance’s governance model: nobody knows who decides, under which membership rules or on what release schedule, and no dated joint roadmaps or deliverables have surfaced.

For Kevin Kirkwood, CISO at Exabeam, without the leading frontier labs the initiative remains “incomplete without broader participation” (Forbes). NVIDIA, for its part, invites governments, industry and researchers to join the alliance’s work.