When hundreds of AI agents being tested by OpenAI broke out of their sandbox and probed Hugging Face systems, the company faced a flood of unfamiliar activity. Its engineers used an open-weight model called GLM 5.2, from the Chinese company Z.ai, to help analyze what had happened.
For Hugging Face CEO Clément Delangue, the episode sharpened a larger argument: AI’s risks may grow as capabilities and resources concentrate in a small number of closed labs, while open models can help more people understand, build with and defend against the technology.
An attack that raised questions
The incident took place in July. At first, Delangue said, the activity did not seem serious enough to become a top priority. Its goal appeared narrow, and the team did not initially consider it highly dangerous or damaging.
What stood out was the scale of the agents’ actions compared with that limited goal. The pattern did not fit what Hugging Face’s team recognized, prompting an investigation. The agents’ speed and the volume of data made the incident difficult to assess using familiar methods.
The open-weight model helped the engineers make sense of that deluge. Delangue sees this as an example of a broader role for open AI: tools that are available for others to examine and run can contribute to understanding and defense, even when the risks come from advanced systems.
Why openness matters to Delangue
Delangue argues that AI naturally tends to concentrate power. Talent, computing capacity and other resources can give a few closed-source companies a growing advantage. He says much of the future compute capacity has already been reserved by a small number of companies.
That concentration matters because AI is becoming a foundation for software and other capabilities. Delangue compares the prospect to a world where only a few companies could write software: in his view, allowing a similarly small group to control AI would place too much influence in too few hands.
He says open-source AI has become a credible alternative to closed-source APIs, but that closed systems still dominate. Giving open AI a stronger chance, he said, requires more resources. That was his explanation for Hugging Face’s acquisition by Nvidia for $12.93 billion in September, despite the company’s commitment to democratizing AI.
Access for research and national sovereignty
Delangue also makes a case for open-weight models as a resource for academia and for understanding AI. He said researchers at Stanford or MIT have no choice but to use Chinese open-weight models because those are the available open options.
He named GLM, DeepSeek, Kimi and Qwen when discussing the need for powerful US open-weight models. In his view, building such models matters for sovereignty as well as research. He sees momentum, but says it is not enough and that more investment is needed.
Some of the strongest non-Chinese open-source models, he said, come from outside the US: Mistral from France, Cohere from Canada, and emerging work in South Korea. He believes the US could do more.
Openness as a way to follow fast-moving AI
Delangue does not dismiss concerns about AI agents or recursive self-improvement. He points out that people already use computers to write code and build tools, and frames the open question as how quickly that process accelerates.
His response is that keeping AI behind closed doors in a few labs does not automatically make it safer. Openness, he argues, gives more people a way to understand and follow developments, and to regulate them if needed. In his view, distributing capability is especially important if AI approaches artificial general intelligence or superintelligence, though he says comparing AI directly with human abilities is not always useful.
He also argues that open models can compete in practical settings. Companies that post-train a model for a specific use case, he said, have reported frontier or better-than-frontier performance at a fraction of the cost. The attack on Hugging Face offers a concrete example of another use: an open-weight model helping engineers investigate an AI-driven incident.
For Delangue, the point is not that openness removes risk. It is that broad access can help people understand the systems being built, respond when something goes wrong and avoid leaving AI’s benefits and control in the hands of only a few companies.