Open-Source Chinese AI Systems Narrow the Gap with Proprietary Cybersecurity Models in Network Penetration Tests

Researchers conducting a controlled experiment have found that two Chinese-developed artificial intelligence models, DeepSeek and GLM, demonstrated network intrusion capabilities that nearly match those of the most advanced cyber‑security model released by OpenAI, a system that remains unavailable to the public.

The test involved prompting each model to identify vulnerabilities and gain access to a series of simulated corporate networks. While OpenAI's internal model achieved the highest success rate, the performance of DeepSeek and GLM was only marginally lower, indicating that open‑source or publicly released AI tools are rapidly acquiring sophisticated hacking techniques.

These findings arrive amid growing debate over the dual‑use nature of powerful language models. Developers of large‑scale AI systems have warned that the same capabilities that enable natural‑language assistance can be repurposed for malicious activities such as phishing, code exploitation, and automated vulnerability discovery.

DeepSeek, a model originating from a Beijing‑based startup, and GLM, produced by a major Chinese research institute, are both marketed for general‑purpose applications and are freely accessible to developers worldwide. Their unexpected proficiency in the simulated attacks raises questions about the adequacy of current oversight mechanisms for open AI releases.

Security experts note that the experiment does not prove the models can independently launch real‑world attacks, but it does suggest that adversaries could fine‑tune publicly available systems for illicit purposes. The gap between publicly released and proprietary models appears to be shrinking, a trend that could accelerate as more organizations contribute training data and compute resources.

In response, several cybersecurity firms are exploring ways to detect AI‑generated attack patterns, while policymakers are considering regulations that would require disclosure of a model’s potential for misuse before it is released. The challenge, however, lies in balancing innovation with safety, as overly restrictive measures could stifle beneficial research.

Future work is expected to expand the scope of testing, incorporating more diverse network configurations and real‑time defenses. By benchmarking AI models against each other, the community hopes to develop standards for responsible deployment and to better understand the evolving threat landscape posed by increasingly capable artificial intelligence.