OpenAI Cyberattack Fallout Spurs Nvidia, Microsoft-Led AI Safety Alliance

"Closed models alone have limits in responding quickly to cyberattacks"

Most open-weight models now come from China

A new AI safety initiative led by major technology companies including Nvidia Corp. and Microsoft Corp. has launched as fallout continues from an incident in which an OpenAI model broke out of control and carried out a cyberattack on US startup Hugging Face.

CNBC and other outlets reported on July 27 that Nvidia announced in a blog post that it, Microsoft, SpaceX and dozens of other US and European technology companies had launched the Open Secure AI Alliance.

In a statement, Nvidia said the alliance would work to address and disclose vulnerabilities using open technology. The recent Hugging Face security incident, it added, underscored the need for cyber defenders to have access to open, state-of-the-art agentic systems to protect themselves.

Last week, Hugging Face was shown to have failed to defend itself despite using leading US defense models. Those models had safeguards that did not distinguish between attackers and defenders. Hugging Face ultimately chose a self-hosted Chinese open-weight defense model that was not bound by those constraints.

Open models can be downloaded, modified and self-hosted. By contrast, AI models built by Anthropic and OpenAI are closed models that can be accessed only through specific infrastructure.

In the US, calls are growing to restrict access to models developed by Chinese AI companies. Lawmakers are mainly pursuing measures to curb the spread of Chinese AI models, which are largely open-weight models.

Chinese companies are known to have built their models through so-called distillation, extracting large volumes of outputs from advanced US AI models to sharply reduce training costs.

Last week, US Treasury Secretary Scott Bessent warned that he would impose sanctions on Chinese companies that carry out piracy attacks targeting American firms.

Chris McGuire, a senior fellow for China and emerging technologies at the Council on Foreign Relations, told CNBC there is a real possibility the US government will impose restrictions on Chinese-made models. Those steps could include banning transactions such as buying tokens through APIs or preventing US cloud companies from hosting Chinese AI models and billing for their use.

He said the debate in Washington would focus on Chinese companies rather than the open-source ecosystem, and on whether to tolerate China's theft of intellectual property. At the same time, concerns are being raised over sanctions on Chinese AI companies because many of the strongest open-source models now come from Chinese firms.

Last week, more than 20 companies including Nvidia, Microsoft, Meta Platforms Inc. and Palantir Technologies Inc. published a letter urging policymakers to avoid "premature regulation of open-weight AI models." The companies said such rules could stifle competition or drive innovation overseas.

Nvidia said the OpenAI incident involving Hugging Face demonstrated a "practical truth." If defenders cannot inspect, adapt and run advanced AI on their own infrastructure, their ability to respond will be limited at the moments when speed matters most.

Kim Jung-a, contributing reporter