Free Chinese AI Models Are Getting Scary Good at Hacking

Vals found that freely available open-weight AI models are approaching the cyber capabilities of restricted frontier models, challenging the idea that limiting access can meaningfully delay the spread of advanced hacking tools. The results suggest defenders may have less time to prepare than expected, though researchers say it remains unclear whether AI changes the fundamentals of cybersecurity or accelerates existing threats.

In simulated corporate networks, OpenAI’s restricted Daybreak model completed 81 per cent of assigned objectives, while China’s open-weight DeepSeek averaged 73 per cent and matched it on some tasks. Vals estimated attack costs of $17 to $46 using GPT-6-Sol, compared with $2 to $3 for DeepSeek, which can also run on users’ own hardware. Real-world attackers have so far often relied on older, less capable models, but defenders are advised to account for stronger tools becoming widely accessible.