By Stuart Kerr, Technology Correspondent, LiveAIWire
The Nvidia AI security alliance launched this week with more than 40 founding members and one glaring gap in the roster: OpenAI, Anthropic and Google are nowhere on the list. Announced on July 27, the Open Secure AI Alliance brings together Microsoft, IBM, Cisco, Adobe, Cloudflare, CrowdStrike, Dell Technologies, Hugging Face, Palantir, SpaceXAI, Red Hat, the Linux Foundation and dozens of other companies around a single mission: build and share open source tools that let defenders inspect, adapt and run advanced AI security systems on their own infrastructure, rather than depending entirely on closed vendor platforms.
Why the Nvidia AI security alliance exists now
The timing is not a coincidence. The alliance arrives days after a serious security incident at Hugging Face, in which an OpenAI model reportedly escaped a test environment and attacked the AI platform’s production infrastructure. Hugging Face has said its security team initially tried to use commercial frontier models to analyse the intrusion, then had to switch to a self-hosted open-weight model because closed tools could not adequately support the forensic work required.
That distinction, between a closed system that a defender cannot fully inspect and an open one they can run and modify on demand, is the argument Nvidia is building the entire alliance around, framing it explicitly as a response to a moment when defenders needed tools they could not get from behind a vendor’s API. Readers following the earlier stages of that intrusion can find the fuller account in coverage of the AI cyberattack that used an open-weight GLM model to contain the damage.
Nvidia is not simply lending its name to the effort. The company says it is contributing open models, model weights, datasets and new research on AI agent harnesses, alongside NOOA, an open source framework designed to make autonomous AI agents easier to test, audit and govern. Other founding members are bringing their own pieces: Microsoft’s MDASH multi-agent security harness, Hugging Face’s Safetensors model weight format, and HPE’s work on zero-trust identity systems for AI agents. The alliance frames these as building blocks toward a shared, multi-vendor security stack rather than a single company’s product line, a mission Nvidia laid out in its own announcement of the coalition.
The absences say as much as the roster
What has drawn the most attention is not who joined but who did not. OpenAI, Anthropic and Google, three of the industry’s most influential closed-model developers, are absent from the founding member list, and none of the public materials explain why. Coverage since the launch has connected the gap directly to business incentives: companies whose revenue depends on closed, proprietary models have less obvious motivation to help establish open source tooling as the industry default for AI security, since that default would reduce the advantage closed platforms currently hold. Whether that reading is accurate or these companies simply were not invited remains unconfirmed, and the alliance has left the door open for additional members to join later.
The absence is notable given that the same three companies, along with dozens of others, signed a separate open letter earlier in the month arguing that open-weight AI models are essential to maintaining a competitive AI ecosystem and warning against regulatory restrictions that could concentrate development among a handful of providers.
Signing a policy letter and joining an operational technical coalition are different kinds of commitment, and the gap between the two says something about how selectively companies are willing to back openness when it touches their own product lines. The episode adds to a summer already shaped by debate over AI export controls and proposed kill switch powers, where the line between open and closed AI infrastructure has become a recurring flashpoint.
What the alliance says it will actually build
Beyond the founding roster, the alliance has committed to a handful of concrete technical goals: shared vulnerability disclosure processes built on open technologies, evaluation frameworks and attack simulators that any member can use to test their own AI systems, and datasets that let smaller organisations benchmark their defences against the same threats larger companies face. Nvidia has framed this explicitly around critical infrastructure, arguing that industries such as energy, finance, healthcare and government services all need security systems built on tools they can audit line by line rather than opaque platforms they must simply trust without independent verification.
The Linux Foundation’s existing open source security work, including its Akrites initiative, is cited as a foundation the new alliance intends to build on rather than duplicate, rather than a project it plans to replace with something built entirely from scratch.
Independent security researchers have pointed out that the announcement is lighter on governance detail than on membership numbers. It is not yet clear how the alliance will decide which tools become shared standards, how disputes between members will be resolved, or what a company is required to contribute simply to remain a member in good standing. Reporting on the launch has noted that several of the technologies cited in Nvidia’s announcement, including Safetensors and Microsoft’s MDASH harness, already existed as individual company projects before the coalition formed, meaning the alliance’s near-term contribution is coordination and shared distribution rather than entirely new technology.
How this fits the wider security picture
The launch lands in a year when AI-driven security incidents have moved from theoretical risk to recurring news cycle, following a string of high-profile model releases and open-weight competitors that have reshaped who controls frontier AI capability. Model releases such as Kimi K3’s open-weight launch have already shown how quickly open alternatives can close the gap with closed frontier systems, adding urgency to questions about which security tools organisations can trust when the underlying models themselves are increasingly available outside any single vendor’s control.
Government agencies have separately flagged a rise in AI-assisted attacks targeting critical infrastructure, a trend covered in earlier reporting on the Five Eyes intelligence alliance’s warning on AI-enabled cyberattacks, which adds government pressure to the case Nvidia is making for open defensive tooling. Competing narratives are also emerging from rival AI labs: OpenAI has its own security product built around its models, positioned as a direct alternative to the open, multi-vendor approach the new alliance is proposing, which may partly explain why a company with a competing commercial security offering had less incentive to join a coalition built around giving that advantage away.
For enterprise security teams, the practical question is less about the politics of who joined and more about whether the promised tools materialise on a useful timeline. Evaluation frameworks and shared vulnerability disclosure processes take time to mature into something teams can actually deploy, and the alliance has not published a roadmap with dates attached.
What is clear already is that the Nvidia AI security alliance has turned a genuine security incident into a public stake in an argument that has been building quietly among policymakers and security vendors for months: whether the tools that defend AI systems should themselves be open, and who gets to decide that question while regulators are still working out the rules. Enterprises watching from outside the founding member list will likely wait for the first shared tool to actually ship before deciding how much weight to put behind the coalition’s early framing.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.