Tests found that the system identified several established vulnerabilities while also working on six categories it was not originally tuned for.
The system, called MOS, adapts a large language model by giving it several specialized networks, each focused on different vulnerability patterns. A routing system selects the networks most relevant to the code being examined, while training data includes explanations checked by experts.
Across four established vulnerability types—reentrancy, timestamp dependency, integer overflow or underflow, and delegatecall—MOS improved average F1 score by 6.38% and accuracy by 4.84% compared with existing methods. In further tests covering access control, price-oracle manipulation, logic errors, missing input validation, slippage, and risky first deposits, it achieved an average F1 score of 83.94% and average accuracy of 87.82%.
What the AI system found
MOS consistently outperformed the comparison methods across the four established vulnerability categories. For integer overflow or underflow, it reached 91.53% accuracy and an 82.35% F1 score, exceeding the best comparison method by 6.88 and 3.60 percentage points, respectively. For delegatecall vulnerabilities, it reached 93.96% accuracy and an 89.52% F1 score, 4.91 and 9.75 percentage points above the best comparison method.
The system also produced explanations for its decisions. In a combined human and AI evaluation, explanations received positive ratings—scores of 3 or 4 on a four-point scale—in 82.96% of cases for correctness, 85.21% for completeness, and 94.58% for conciseness. In tests on six additional vulnerability categories, MOS outperformed all the comparison methods and achieved the average scores reported above.
Why broader detection matters
Smart-contract vulnerabilities can put blockchain-based systems and digital assets at risk. A detection system that identifies more than one kind of flaw, explains its decisions, and remains effective on vulnerability categories added later could be more useful for examining varied code than systems built around a fixed set of rules.
The results also suggest that MOS is not limited to the four vulnerability types used in its main evaluation. Its performance on six additional categories indicates that the approach can extend to vulnerability types beyond those for which it was initially specialized, although the experiments do not establish how it would perform in every real-world setting.
Evidence and caveats
The evidence comes from experiments comparing MOS with existing vulnerability-detection methods, plus human and AI assessments of its explanations. The reported results are evaluations of the system on selected vulnerability categories rather than a demonstration of deployment across all smart contracts.
The researchers examined 10 representative misclassified cases for each of four categories. They found both unnecessary warnings and missed vulnerabilities: MOS tended to flag some secure patterns and struggled with complex cases. The system also requires substantial computing resources: training required at least two NVIDIA GeForce RTX H800 GPUs with 80 gigabytes of memory each, while the full model contains 18 billion parameters even though about 5 billion are activated during inference. The researchers further note that different prompt designs might improve performance.
// Source
ACM Transactions on Software Engineering and Methodology · 2026 · DOI: 10.1145/3859369
Authors: Hang Yuan, Lei Yu, Zhirong Huang, Jingyuan Zhang, Junyi Lu, Shiqiang Cheng, Fengjun Zhang, Jiajia Ma, Chun Zuo
Institutions: Chinese Academy of Sciences, University of Chinese Academy of Sciences, Institute of Software