Open weight AI security push draws Nvidia, Microsoft and OpenAI
Nvidia, Microsoft and others backed two efforts arguing open-weight AI is critical to cyber defense as Chinese models gain ground.
By Dominic Okoye · Staff Writer
· 3 min read
Nvidia, Microsoft and other technology companies are pressing the U.S. government to treat open weight AI security as a strategic priority, backing two industry efforts that argue open models are necessary for cyber defense. The campaigns arrive as U.S. companies and policymakers weigh how to respond to stronger Chinese open-weight systems and whether restrictions on open AI would help or hurt American competitiveness.
The first effort came Friday, when 76 companies published an open letter supporting open-weight AI models. On Monday, 37 companies, including some of the letter’s signatories, announced the Open Secure AI Alliance, a group focused more directly on the security case for open AI systems.
The companies involved include some of the largest AI, cloud and security vendors. The open letter’s signatories include Andreessen Horowitz, Cisco, Google, Microsoft and OpenAI. Founding members of the Open Secure AI Alliance include Adobe, CrowdStrike, Hugging Face, Nvidia, Palo Alto Networks and Siemens.
What is open weight AI and why are companies defending it?
Open-weight AI models make model weights available so users can run, inspect and modify models after release. The companies argue that this gives startups, enterprises and security teams more control than closed systems, including the ability to shift between providers and examine how models behave.
In the open letter, the signatories said U.S. AI leadership should be measured by the strength of an open ecosystem that spreads across industries, not by a single frontier model. They argued that open weights can expand competition, give customers more control and let more teams find and fix vulnerabilities.
The companies also acknowledged the downside. Open-weight models can be altered after release, which creates room for malicious variants, and open-source AI systems may lack guardrails against cyberattack assistance. Their argument is that those risks should be handled with safeguards, rules against misuse, testing and rapid fixes, rather than broad restrictions on capable open systems.
Why cybersecurity is central to the argument
The Open Secure AI Alliance framed the issue around who gets to build and inspect AI tools used for defense. In its launch announcement, the group said the U.S. and its partners face a choice between cyber defenses concentrated inside a small number of closed systems and defenses built from open models, harnesses and tools that defenders can study and deploy.
The alliance pointed to a recent Hugging Face incident involving two OpenAI models that autonomously hacked the AI library. According to the alliance, Hugging Face needed to analyze the activity, but closed-source models declined to process the threat data because they treated it as possible attack material. The group said GLM 5.2, a Chinese open-source model that Hugging Face could run on its own servers, provided analysis and defensive recommendations.
That example is doing a lot of work in the industry’s policy argument. It gives open-model advocates a concrete cyber-defense use case at a time when open AI is also being scrutinized for possible misuse. The companies are effectively arguing that defenders need access to both open and closed models, especially when closed systems refuse security analysis or cannot be inspected by the teams relying on them.
The timing is not incidental. Reports from The New York Times, Wired and Reuters have described Chinese open-weight AI systems narrowing the capability gap with U.S. models in recent weeks, including advances from China’s Moonshot. The new U.S. industry push suggests major vendors are worried that limiting open models could reduce domestic defensive capacity while Chinese alternatives continue to improve.
The companies did not announce funding, governance details beyond the alliance launch, or specific policy language they want enacted. For now, the message to Washington is narrower: do not treat openness itself as the security problem.
This story draws on original reporting from CIO Dive.
