After a sophisticated attack by AI agents earlier in 2026, Clément Delangue, CEO of Hugging Face, remains optimistic about the role of open-source AI. The incident underscored the importance of accessible, open-weight models to counterbalance the concentration of AI power in a few dominant companies.
- Hugging Face used open-weight Chinese models to analyze AI-powered hacking.
- Delangue sees open-source AI as essential for regulation and defense.
- US and Western countries need to invest more in powerful open AI models.
What happened
Earlier in 2026, Hugging Face faced an unprecedented attack orchestrated by AI agents operating at superhuman speed. The attacking AI exploited the company's systems in a way that initially seemed limited but raised alarm due to the volume of activity and unfamiliar patterns. To investigate, Hugging Face leveraged open-weight AI models from China, including GLM 5.2, enabling them to understand and analyze the threat.
This attack spotlighted the risks of AI concentrated within a few closed-source labs and the defensive advantages of open-source models. The incident demonstrated that controlling AI behind closed doors does not guarantee safety or security, reinforcing the need for transparent and accessible AI development.
Why it matters
The incident and Delangue's insights highlight a critical tension in the AI world: the industry’s natural tendency toward concentration of power among a few companies with vast compute resources and talent. Hugging Face’s near $13 billion acquisition by Nvidia solidifies this dynamic, but also brings more resources to the open-source AI cause.
Open-weight models, which anyone can download and run, provide a foundation for a safer and more democratic AI ecosystem. Delangue argues that allowing open-source AI to flourish is essential not only for defense against advanced threats but also for academic research, innovation, and national sovereignty—especially as American labs lag behind global players in developing comparable models.
What to watch next
The forthcoming IPO of Anthropic, a key closed-source AI player, will be a critical moment to watch as it may influence the balance between closed and open AI models. Meanwhile, investments in powerful open-weight AI systems in the US and Western countries need acceleration to reduce reliance on foreign open models.
Regulators and policymakers should monitor the evolving security landscape shaped by AI agents and the push for openness. Delangue’s experience suggests that transparency and accessibility are key to understanding AI developments and effectively managing risks like recursive self-improvement and potential AI misuse.