A recent experiment tested how well AI models could break into simulated networks. The results showed that free Chinese models are catching up quickly in offensive cyber capabilities. Two models, DeepSeek and GLM, performed surprisingly well.
The test measured the hacking skills of several AI systems against simulated network targets. DeepSeek and GLM demonstrated abilities that came close to OpenAI’s most advanced cyber model. That model is not available to the public.
Researchers designed the experiment to compare models from different developers. The Chinese models are freely accessible, while OpenAI restricts its top cyber model. Despite this gap in access, the performance difference was smaller than expected.
DeepSeek and GLM showed competence in tasks like reconnaissance and exploit generation. These are core steps in many real-world cyberattacks. Their results suggest open models are becoming more capable in security-sensitive domains.
The findings raise questions about how freely available AI tools could be misused. Lower barriers to advanced hacking capabilities may increase risks for organizations. Security teams may need to rethink their defenses against AI-assisted attacks.
OpenAI’s advanced cyber model remains closed to the public for safety reasons. The company has cited concerns about misuse in offensive operations. The new results show that open alternatives may offer similar capabilities without those safeguards.
The experiment adds to a growing debate over releasing powerful AI models. Some argue that transparency and open access drive innovation. Others warn that unrestricted models can lower the cost of launching cyberattacks.
For now, the gap between open and closed cyber models appears to be narrowing. DeepSeek and GLM show that free tools can deliver results once limited to top-tier systems. That trend is likely to shape how defenders and attackers use AI in the months ahead.
