TianGong Cyber Preview: 27B Security LLM Beats GPT-5.5 Cyber on Real-World Exploits
Ant Group's TianGong Cyber Preview, a 27B-parameter security LLM built on Qwen, achieves 90.6% vulnerability trigger rate on Cybergym, leads on CyberGym-E2E under $10 budget, and discovers real-world Linux kernel and Solana Pay vulnerabilities, demonstrating high security intelligence-efficiency ratio for practical deployment.
Ant Group's AI Security Lab, in collaboration with Tsinghua University and Zhejiang University, has released the TianGong Cyber model Preview — a cybersecurity-focused large language model trained on the Qwen 27B base. The model aims to deliver frontier security capabilities with a high "security intelligence-efficiency ratio" (安全智效比), addressing the industry pain point of high inference costs and closed-model monopolies.
Training Methodology: Agentic Scaling Pipeline
The team built an efficient Agentic post-training pipeline integrating data scaling, environment scaling, and training scaling. Real-world offensive and defensive tasks were transformed into large-scale, high-quality, verifiable training environments. Expert insights from Ant's Light-Year Lab (winner of Tianfu Cup, GeekPWN, Pwnie Awards, top Apple vulnerability acknowledgments) and academic partners were fused into the process, driving coordinated Agentic SFT and RL under a unified verification standard.
Benchmark Results
Cybergym: Vulnerability Discovery & Exploitation
On Cybergym, the model must autonomously perform long-horizon vulnerability analysis, discovery, and verification across 1,507 target programs in 188 complex software systems. TianGong Cyber achieved a 90.6% vulnerability trigger success rate , outperforming Gemini-3.8-Flash Cyber and GPT-5.5 Cyber.
CyberGym-E2E: End-to-End Offensive/Defensive Tasks
This benchmark covers 920 real vulnerabilities in 139 open-source projects, requiring analysis, PoC triggering, patching, and regression testing. Under a $10 per-task budget , TianGong Cyber scored 78.5% , leading other models under the same budget constraint, though it still trails Opus 5 without budget limits.
EVMBench: Smart Contract Security
EVMBench tests off-chain auditing (Detect), patching (Patch), and on-chain exploitation (Exploit). TianGong Cyber topped the weighted overall ranking, with Detect and Exploit capabilities at the leading level , validating its transferability to low-resource Web3 security scenarios.
Real-World Validation
Linux Kernel UFS Vulnerability
During a code audit of the Linux kernel UFS file system, the model identified an out-of-bounds read risk caused by missing disk metadata validation in the mount and resource allocation paths. The vulnerability was dynamically verified and reported, demonstrating capability in critical infrastructure software auditing.
Solana Pay Protocol Audit
Jointly with Zhejiang University's Blockchain and Data Security Lab, the model audited Solana blockchain code and discovered a payment interface verification bypass in the Solana Pay Machine Payments Protocol. The risk has been fixed by the Solana Foundation, showcasing Web3 and novel payment system security skills.
Reverse Engineering of Obfuscated Application
In a pure static, local environment, TianGong Cyber performed reverse analysis of a well-known application package, preliminarily restoring program logic and data flow under hardening and obfuscation, confirming closed-source binary analysis ability.
Limitations & Future Work
The authors acknowledge that parameter count remains a key determinant of intelligence ceiling. The 27B model exhibits some capability gaps in complex scenarios; future work will explore different parameter scales to push the security intelligence-efficiency frontier.
Application Scenarios
The model targets diverse security tasks:
Vulnerability tasks: discovery, exploitation, patching
Reverse engineering: black-box analysis, protocol restoration, REA
Offensive/defensive: penetration testing, intrusion detection, fraud opposition, data security
Analysis: traffic analysis, network protocol analysis
Preview Access
Free trial applications and open-sourcing are planned. Interested security teams, open-source maintainers, and industry partners can contact [email protected] for collaboration.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
