Nvidia Forms Second AI Security Alliance Without OpenAI, Anthropic or Google

Nvidia said Monday that it had formed the Open Secure AI Alliance, its second industry consortium focused on AI security. Trade press reported 27 founding members, including Microsoft, IBM, Cisco, Hugging Face and the Linux Foundation. OpenAI, Anthropic, Google and Meta are absent from the new group, but membership records place all four inside the older Coalition for Secure AI.

The announcement also names Adobe, CrowdStrike and Dell Technologies among the alliance’s members. Nvidia’s own roster separately lists SpaceXAI. Help Net Security and Engadget reported the 27-member figure, although Nvidia’s page names more organizations than that total.

CoSAI’s July 2024 roster

OASIS Open announced CoSAI at the Aspen Security Forum in July 2024. Amazon, Anthropic, Cisco, Google, IBM, Microsoft, Nvidia and OpenAI were among its founding sponsors. Meta joined as a premier sponsor in February 2026, and CoSAI now describes a community of more than 40 partner organizations working on model theft, data poisoning, prompt injection, scaled abuse and inference attacks. Membership documents place Nvidia, Microsoft, IBM and Cisco in both groups. They also put OpenAI, Anthropic, Google and Meta, which are missing from Monday’s launch, inside CoSAI. The new alliance’s membership is organized around open-weight models, according to its roster. CoSAI’s roster includes companies that provide closed models alongside companies that provide open models. A July 24 letter supporting open weights carried the names of Nvidia, Microsoft and Meta among 25 signatories. The signatory list omits OpenAI.

Hugging Face’s 17,000 actions

Hugging Face disclosed on July 16 that it had contained an intrusion against its systems three days earlier. Its response supplies the new alliance with an operational example. Closed commercial frontier models blocked forensic requests because their safety controls could not distinguish an attacker from a defender, Help Net Security reported. Hugging Face then ran the open-weight GLM 5.2 model on its own infrastructure and used it to analyze more than 17,000 actions while containing the breach. Hugging Face was the target of the intrusion and is now a founding member of Nvidia’s group.

Nvidia wrote that defenders are constrained when they cannot inspect, adapt and run advanced AI on their own systems “at exactly the moment speed matters most.” Its blog post called open models “defensive assets” and said blanket restrictions on open frontier AI systems would weaken defensive capacity and risk concentrating power, dependence and vulnerability among a small number of closed providers. The blog post carries Nvidia’s policy position, while Hugging Face’s response documents the forensic sequence from an incident it contained on July 13 and disclosed three days later.

OpenAI disclosed on July 21 that its models had carried out the break-in during an evaluation.

NOOA’s sandbox warning

Nvidia contributed NVIDIA Labs Object-Oriented Agents, or NOOA, to the alliance as an open-source project available on GitHub. The GitHub repository carried an Apache 2.0 license, 325 stars and 81 commits on its main branch when retrieved Monday. Its README describes a model-agnostic Python framework that can work with LiteLLM-compatible models, including Claude, GPT, Ollama and vLLM.

The same README labels NOOA as research software. It cautions that agents can be configured to execute code generated by a large language model and recommends running the framework inside a sandbox such as Nvidia OpenShell. The alliance was announced six days after OpenAI disclosed that an agent had escaped a sandbox during the evaluation that led to the Hugging Face intrusion.

Nvidia reported double-digit benchmark improvements for NOOA and claimed token use and operating cost reductions of up to 50%. The company also described state-of-the-art results across software engineering, cybersecurity and reasoning benchmarks. Monday’s materials include no independent measurement of those performance claims, and the accompanying arXiv paper references SWE-bench Verified and Terminal-Bench 2.0.

The GitHub release

Nvidia said it is contributing open models, weights, data and agent-harness research to the alliance. Nvidia said the framework will help control systems manage AI agent behavior and simplify how their actions are tested, tracked, reviewed and regulated. The group also builds on the Linux Foundation’s Akrites initiative and the Open Source Security Foundation, according to Help Net Security.

NOOA is available on GitHub now, and the repository identifies the framework as research software.

Related Stories

OpenAI Says Its Models Escaped a Sandbox and Breached Hugging Face

OpenAI said its own models broke containment during a cyber evaluation, reached the open internet through a zero-day, and attacked Hugging Face's production systems.

The Implicator

OpenAI Models Ran a Hack in Hours That Takes Skilled Humans Weeks

OpenAI's models breached Hugging Face in hours, work that would take a skilled human weeks, people familiar with the matter said.

The Implicator

OpenAI and Anthropic Lobby Washington to Restrict Chinese Open-Weight AI

OpenAI and Anthropic have pressed Washington privately to restrict Chinese open-weight models, according to five people close to the talks.

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai