Nvidia AI safety alliance targets open-model cyber defense

Nvidia is leading 36 other companies in a Nvidia AI safety alliance aimed at strengthening cybersecurity tools built around open artificial intelligence models. The Open Secure AI Alliance was announced after OpenAI said its models were involved in a cyber incident affecting Hugging Face during cyber capabilities testing.

Nvidia described the group in a Monday blog post as a coalition focused on open technologies that can help defenders find and fix vulnerabilities. Participants named by Nvidia include Capital One, CrowdStrike, DoorDash, Microsoft, IBM and SpaceX, among others.

The initiative points to a widening debate over how governments and companies should secure AI systems as advanced models become part of cyber defense and testing. Nvidia said defenders may be limited if they cannot examine AI tools, adapt them to their own needs or operate them on infrastructure they control, particularly when rapid response is required.

What is the Open Secure AI Alliance?

The Open Secure AI Alliance is an industry group organized around open AI models, tools and evaluation methods for cybersecurity. Its stated goal is to support a shared defensive base that companies and public-sector bodies can inspect, modify and deploy across systems from multiple vendors.

Open AI models make their design or operating details more accessible than closed systems, allowing security teams to study and customize them. Nvidia said that approach can broaden access to defensive capabilities, give defenders more visibility, protect data by allowing local deployment and complement closed frontier models with controls tailored to specific organizations or jurisdictions.

Nvidia also acknowledged that open models carry risks. The company said such models, like other technologies, can be misused by actors seeking to weaken safeguards or reshape capabilities for cyberattacks.

The alliance is calling for investment in shared open infrastructure for AI defense, according to Nvidia. The company listed datasets, evaluation frameworks, attack simulators and red-team tools as examples of resources that could be developed collectively, comparing the proposed effort to earlier investment in open source software.

Nvidia said contributors are working on an open defense stack for AI agents. The examples it cited include Microsoft’s MDASH and SpaceXAI’s open sourcing of the Grok Build coding agent.

The announcement followed OpenAI’s statement last week that a cybersecurity incident reported earlier in the month by Hugging Face was caused by OpenAI models during cyber capabilities testing. OpenAI said on July 21 that it viewed the incident as unprecedented and involving state-of-the-art cyber capabilities, and that it was responding on that basis.

The episode has given new urgency to questions about whether AI security should depend mostly on closed systems controlled by a small number of providers or also include open tools that defenders can independently examine. Nvidia’s position, as set out in its blog post, is that both closed and open AI models are needed, with open systems playing a specific role in cybersecurity.

This story draws on original reporting from PYMNTS.