When OpenAI's (OPAI.PVT) research models went rogue and hacked open-source AI site Hugging Face over the summer, cofounder and chief science officer Thomas Wolf says the company couldn't rely on closed models, such as those from OpenAI or Anthropic, to defend itself.
Instead, he told Deirdre Bosa, the company had to turn to an open-weight AI model developed by Chinese company Z.ai to determine what was happening during the attack and cut it off.
The episode turned on its head the thinking that open models were somehow dangerous.
"Traditionally, let's say last year, people had this view that open-weight models might be dangerous, and actually they turned out in our case to be our savior," Wolf said.
Open-weight models are models that developers make available free for users to download, customize, and run on their own systems.
The companies that develop them, including China's Alibaba (BABA), DeepSeek, Z.AI, and others, make money by charging customers who lack the necessary hardware to run the models to access them on their own computers.
Open-weight models, especially those from China, have quickly gained in popularity among developers and enterprises.
The software is generally less expensive to run than leading closed models, and can be tuned to a customer's needs.
When OpenAI's models attacked Hugging Face, Wolf says the company couldn't use closed-weight models because their built-in guardrails wouldn't allow the firm to analyze the relevant hacking data.
But the open-weight model from Z.ai, like most open-weight models, allowed Hugging Face to adjust its guardrails, helping stop the attack.
"I think what we saw in the whole summer was people realizing that …everyone can't rely on these two labs [OpenAI and Anthropic] to protect themselves," Wolf said.
"So we need a lot more tools to protect ourselves, and open-weight models are gonna be one of the major…tools."
Email Daniel Howley at dhowley@yahoofinance.com*. Follow him on X at* @DanielHowley or reach out on Signal or Telegram.