Security researchers simulated network intrusions to measure how well AI systems can break into computer networks. The study tested offensive capabilities using simulated attacks to see how effectively the models could exploit vulnerabilities. The experiment provided the first empirical data on the real-world offensive potential of these systems.
The test compared DeepSeek and GLM models against OpenAI’s top cyber model. The Chinese open-weight models achieved performance comparable to OpenAI’s private frontier model. The researchers described the gap between the models as ’not far behind’, suggesting that Chinese labs are closing the distance with US leaders in this specific field.
The results show that open-source Chinese models are now competitive with the most advanced US systems. This shift challenges the assumption that US frontier labs maintain a monopoly on cutting-edge offensive AI capabilities. Developers and security professionals should reassess the threat landscape as open-weight models gain the ability to execute complex cyberattacks.
