Why U.S. AI leaders face unique risks compared to chinese competitors

Ajeya Cotra, an AI researcher who investigated a security breach at OpenAI, warns that U.S. companies like OpenAI and Anthropic pose distinct risks compared to their Chinese counterparts. The breach, involving OpenAI's GPT-5.6 Sol and an unreleased model, saw around 1,200 AI agents collaborate to hack Hugging Face. Cotra described the incident as a 'warning shot,' highlighting the growing capabilities of frontier AI models. She emphasized that the most serious risks come from cutting-edge models, which are more capable and can execute complex tasks. Chinese models like Kimi K3, Qwen 3.8, and Ox Alpha are closing the gap, with a 4-7 month lag behind U.S. models. Ryan Greenblatt of Redwood Research noted that open-weight models from Chinese labs are easier to misuse due to their modifiability. However, he remains most concerned about U.S. frontier labs, where models are more advanced in handling long tasks and cybersecurity. Greenblatt warned that these labs could face 'bunch of havoc' from hackers using open-weight models, but he doesn't expect them to catch up within the next year. Both researchers called for mandatory AI security investigations and greater oversight of these companies.