OpenAIWARN: Chinese DeepSeek Uses Distillation to Access U.S. AI Models

OpenAI warned U.S. congressional leaders on February 13, 2026, that Chinese AI startup DeepSeek has sought to access U.S. AI model outputs through "distillation" techniques to train its own models. The memo details that DeepSeek employees have developed methods to circumvent OpenAI access restrictions and accessed models via obfuscated third-party routers. OpenAI also found code being developed to automate access to U.S. AI models for distillation training. DeepSeek’s models, introduced at the start of last year, outperformed some U.S. offerings with far lower compute requirements, intensifying Washington’s concerns over China’s rapid AI advancement despite restrictions on high-performance computing chips.