Free Chinese AI Models Are Getting Scary Good at Hacking

According to new research from Vals, open-weight artificial intelligence models are now nearly as capable of carrying out complex, multi-step network hacking instructions as the best closed models. Researchers found that while OpenAI's advanced proprietary model Daybreak completed the highest percentage of tasks, the Chinese open-weight model DeepSeek followed closely behind at a significantly lower cost. The findings suggest that efforts by US frontier labs to limit model access in order to slow the proliferation of advanced cyber capabilities may be losing their effectiveness. Open article →