Chinese AI Models Match OpenAI's Top Cyber Capability in Hacking Tests

In a controlled experiment, DeepSeek and GLM demonstrated hacking skills nearly equivalent to OpenAI's most advanced (non-public) cyber model when tasked with breaching simulated networks. This suggests that open-source Chinese AI models are closing the gap in offensive cyber capabilities, potentially lowering the barrier for cyber operations. The findings are based on a single experiment and may not reflect real-world performance, but they signal a shift in the AI-cyber landscape.

AI disclosure: Model-written summary of the source, not a verified fact. Check the sources. How Mosaic works