Alibaba's XekRung Tops CyberGym Leaderboard at 88.9%
Alibaba's security-focused LLM XekRung-1.5-27B-Preview took first place on UC Berkeley's CyberGym vulnerability-reproduction benchmark with an 88.9% success rate. Fine-tuned from Qwen3.8-27B, which scored 54.51%, it beat Gemini 3.8 Flash Cyber (86.3%) and GPT-5.5-Cyber (85.6%). Weights remain unreleased.
- XekRung-1.5-27B-Preview scored 88.9% on CyberGym; base Qwen3.8-27B scored 54.51%
- CyberGym holds 1,507 real vulnerabilities from 188 OSS-Fuzz open-source projects
- It leads Gemini 3.8 Flash Cyber (86.3%), GPT-5.5-Cyber (85.6%) and GLM-5.3 (84.5%)
- Model weights are not released and no timeline for outside access was given
Read next
AI