Nvidia leads push for open AI cyber tools after Hugging Face hack

Nvidia and top tech companies have assembled a group calling for better support for open AI models to fend off cyberattacks in the wake of OpenAI’s rogue agent hacking Hugging Face.

The Open Secure AI Alliance, unveiled Monday, comprises dozens of companies, including Microsoft, SpaceXAI, IBM, Palantir, Databricks, Dell, and Hugging Face itself, as well as the Linux Foundation and startups such as Cognition and Thinking Machines Lab.

In a blog post announcing the group, Nvidia called for regulators to recognize open models and security tooling as defensive assets rather than liabilities.

Blanket restrictions would “weaken defensive capacity,” it warned, and concentrate dependence on a few providers.

Critics view open models as a liability because their safeguards can be stripped out and their capabilities repurposed for attacks, a risk the post acknowledged but said was not unique to open systems.

Nvidia’s post, which lists the companies as inaugural partners, also calls on governments and companies to fund shared open datasets, evaluation frameworks, and red-teaming tools.

The group frames closed and open models as complementary. “Attackers have frontier AI.

Defenders need a frontier AI ecosystem—the best open and closed models, force-multiplied by a global community,” Nvidia CEO Jensen Huang wrote in a Monday X post. “During the Hugging Face incident, closed AI blocked essential forensics.

An open-weight frontier model helped contain the intrusion.

That’s why we created the Open Secure AI Alliance,” he added.

The announcement comes after Hugging Face said on July 16 that an AI agent had broken into its systems and stolen an access key.

It said it used the key to reach deeper into the company’s network.

Five days later, OpenAI said that its models were responsible: GPT-5.6 Sol and a more capable pre-release system, both running with safety refusals dialed down for an internal test.

Hugging Face first tried to investigate the breach using commercial AI services, but analyzing the intrusion required submitting the attacker’s own code.

Closed tools, “unable to distinguish attackers from defenders,” blocked the analysis, Nvidia said.

The safety filters designed to stop hackers ended up hindering the hacked company instead.

Instead, Hugging Face ran an open model, Z.ai’s GLM 5.2, on its own servers to review more than 17,000 actions.

Defenders “need open, frontier agentic systems for self-defense,” the Nvidia blog argued, or their response is constrained “at exactly the moment speed matters most.” Nvidia does not name OpenAI in the post, referring only to “closed AI tools.” OpenAI is also absent from the membership list, along with Anthropic, Google, Meta, and Amazon.

Z.ai, the Chinese developer of the model used by Hugging Face, is also not included.