Open‑Source Chinese AI Models Close In on Proprietary Cybersecurity System in Penetration Testing

A controlled study by researchers revealed that the Chinese‑built AI systems DeepSeek and GLM showed network‑breaching abilities that almost equal those of OpenAI’s top‑tier cybersecurity model, which has not been released to the public.

In the experiment, each model was asked to spot flaws and infiltrate a set of mock corporate networks. OpenAI’s internal system recorded the best success rate, but DeepSeek and GLM trailed only slightly, suggesting that publicly available or open‑source AI is quickly mastering advanced hacking methods.

The results come as the conversation around the dual‑use potential of potent language models intensifies. Creators of large‑scale AI have cautioned that the very functions that power natural‑language help can also be turned toward malicious ends, including phishing, code exploits, and automated discovery of vulnerabilities.

DeepSeek, developed by a Beijing startup, and GLM, from a prominent Chinese research institute, are both positioned as general‑purpose tools and are openly available to developers globally. Their surprising skill in the simulated breaches prompts scrutiny of whether existing oversight for open AI releases is sufficient.

Security analysts point out that while the study does not demonstrate that the models can autonomously conduct real‑world attacks, it does imply that malicious actors might fine‑tune publicly accessible systems for illicit use. The distance between open‑source and proprietary models seems to be narrowing, a pattern likely to quicken as additional entities supply training data and computing power.

Consequently, a number of cybersecurity companies are investigating techniques to spot AI‑crafted attack signatures, and legislators are debating rules that would mandate disclosure of a model’s misuse risk prior to release. The difficulty remains striking a balance between fostering innovation and ensuring safety, since excessive restrictions might hamper valuable research.

Upcoming research aims to broaden the testing parameters, adding varied network setups and live defenses. By pitting AI models against one another, the community seeks to forge standards for responsible deployment and gain clearer insight into the shifting threat landscape driven by ever‑more capable artificial intelligence.