DeepSeek Accused of Copying Top U.S. AI Models Since Late 2024

On Sept. 8, 2026, U.S. cybersecurity agency CISA said in advisory AA26-251A that DeepSeek has run an organized campaign since at least late 2024 to distill — use outputs from — leading U.S. AI models to train its own. I’m skeptical of cost forecasts: if copying at scale works, big training budgets offer less protection.