Alibaba's AI Security Model Tops CyberGym Leaderboard, First Chinese Model to Rank First

On September 28, Alibaba's Fuyuan Xirang large model (XekRung-27B) topped the CyberGym model leaderboard with an 88.9% success rate, becoming the first Chinese model to take first place in the internationally recognized AI security benchmark.

Built on Alibaba's self-developed Qwen3.8-27B multimodal large model, XekRung-27B underwent security-specific training and extreme compression. Despite having only 27B parameters—roughly 1/27 to 1/370 of comparable models—it achieved the highest success rate while significantly reducing computing costs for security vulnerability detection.

CyberGym, operated by a research team at UC Berkeley, is a benchmark framework covering 1,507 real security vulnerabilities across 188 software projects. The leaderboard is split into model and agent categories. Alibaba topped the model category, which uses only a generic harness without external tools or knowledge bases, reflecting pure security intelligence.

Models from Google, OpenAI, Zhipu, and DeepSeek ranked second through fifth with success rates of 86.3%, 85.6%, 84.5%, and 83.3%, respectively. Huang Longtao, head of Alibaba's Security AGI Lab, said AI has significantly improved detection rates. He added that Fuyuan Xirang will continue evolving in areas like adversarial intelligence, self-evolution, and agentic capabilities, and that Alibaba plans to make its security technologies available to society through multiple channels.