MAS-SZZ : Multi-Agentic SZZ Algorithm for Vulnerability-Inducing Commit Identification CCS: Software and its engineering Software maintenance tools
Sicong Cao † , Jinxuan Xu † , Le Yu † ♣ , Jing Yang ‡ , Xingwei Lin ‡ , Linlin Zhu § , Fu Xiao † · 2026
Read MAS-SZZ : Multi-Agentic SZZ Algorithm for Vulnerability-Inducing Commit Identification CCS: Software and its engineering Software maintenance tools by Sicong Cao † , Jinxuan Xu † , Le Yu † ♣ , Jing Yang ‡ , Xingwei Lin ‡ , Linlin Zhu § , Fu Xiao † (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-24398.txt.
Introduction
Cao and six colleagues present MAS-SZZ, a multi-agent method for faulty commits. The SZZ algorithm traces history to commits touching vulnerable code, but variants fail through wrong anchors and weak backtracking. Given a CVE description and fixing commit, MAS-SZZ sums up the vulnerability's root cause. Step-forward prompting pinpoints lines from hunk intent. Those lines anchor a backward trace to the commit that introduced the flaw. Tests show MAS-SZZ beats baselines across datasets and languages, with F1 gains of about 65 percent over the best method.
Catalog record
| CREATOR | Sicong Cao † , Jinxuan Xu † , Le Yu † ♣ , Jing Yang ‡ , Xingwei Lin ‡ , Linlin Zhu § , Fu Xiao † |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 26833 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.24398
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
MAS-SZZ : Multi-Agentic SZZ Algorithm for Vulnerability-Inducing Commit Identification CCS: Software and its engineering Software maintenance tools Sicong Cao † , Jinxuan Xu † , Le Yu † ♣ , Jing Yang ‡ , Xingwei Lin ‡ , Linlin Zhu § , Fu Xiao † Abstract . Accurate vulnerability-inducing commit identification serves as a foundation for a series of software security tasks, such as vulnerability detection and affected version analysis. A straightforward solution is the SZZ algorithm, which traces back through the code history to identify the earliest commit that modify the vulnerable code. Unfortunately, neither the customized V-SZZ nor state-of-the-art LLM4SZZ perform satisfactorily due to the incorrect anchor selection and inadequate backtracking capability, making them far beyond a reliable usage in practice. To overcome these challenges, we propose a multi-agentic SZZ algorithm, named MAS-SZZ , that facilitates the identification of vulnerability-inducing commits through collaboration among agents. Specifically, given a CVE description and its corresponding fixing commit, MAS-SZZ summarizes the root cause of the vulnerability and employs a structured step-forward prompting strategy to localize vulnerability-related statements based on the change intent of each patch hunk. These vulnerable statements serve as anchors from which MAS-SZZ autonomously traces backward through the repository’s history to find the commit that first introduced the vulnerability. Extensive experiments show that MAS-SZZ outperforms the state-of-the-art baselines across datasets and programming languages, achieving F1-score gains of up to 65.22% over the best-performing SZZ algorithm. 1. Introduction Unpatched software versions with 1-day vulnerabilities pose severe threats to end users, as adversaries could learn from released patches to find and exploit them, incurring massive remediation and operational costs ( 20 ) . Therefore, it is essential to determine which versions are affected by a known vulnerability, underpinning a wide range of downstream tasks such as vulnerability detection ( 15 ) and severity assessment ( 8 ) . A straightforward solution is the SZZ algorithm ( 13 ; 7 ; 4 ; 3 ) , which identifies the Vulnerability-Inducing Commits (VICs) that introduced the security-critical bugs, from the Vulnerability-Fixing Commits (VFCs) . The pioneering work V-SZZ ( 1 ) leverages the line mapping algorithms to identify the earliest commit that modified the vulnerable lines as VICs. Despite receiving great attention from academics and practitioners, neither the foundational SZZ algorithm nor its improved variants perform satisfactorily. Even though the state-of-the-art approach LLM4SZZ ( 16 ) achieves an accuracy of only 40.7% on the manually curated dataset ( 2 ) , which is far beyond a reliable usage in practice. Such poor ability stems from two major challenges : Challenge 1: Incorrect anchor selection. Based on the core assumption that only deleted or modified lines cause vulnerabilities, most SZZ algorithms adopt simple diff-based heuristics to select anchor statements for backtracking. Nonetheless, due to the existence of composite ( 5 ) and ghost ( 11 ) commits, they often fail to localize the true root cause line, resulting in high false-positive rate. Although LLM4SZZ tries to address this issue using Large Language Model (LLM) for contextual filtering, it achieves limited success as vulnerability-inducing statements are sometimes hidden in the unmodified context. Challenge 2: Inadequate backtracking capability. Standard SZZ algorithms apply one-step backtracking, which often misses earlier commits that introduced the vulnerability. V-SZZ is the first to introduce the multiple backtracking capability, which performs consecutive tracing until it identifies a commit as a stopping point. However, it still suffers from the over-traced or under-traced issue because of the use of similarity-based heuristics. Our work. In this paper, we present MAS-SZZ , a novel multi-agentic SZZ algorithm for VIC identification. Specifically, given a CVE description and its corresponding VFC, MAS-SZZ first summarizes the root cause of the vulnerability, accompanied by verified evidence points that trace each claim back to the supplied inputs. Then, MAS-SZZ employs a structured step-forward prompting strategy to infer the change intent of each patch hunk, and identifies the root cause lines from vulnerability-related hunks for backtracking, addressing Challenge 1 . Third, starting from an anchor statement, MAS-SZZ iteratively backtracking the commit history guided by the root cause, and dynamically determines the earliest vulnerability-contained commit as the VIC, resolving Challenge 2 . Evaluation. We implement a prototype system of MAS-SZZ , and conduct comparative experiments with seven representative SZZ algorithms ( 13 ; 7 ; 3 ; 1 ; 4 ; 16 ) on the two well-regarded benchmark datasets ( 1 ; 6 ) . Experimental results show that MAS-SZZ outperforms the state-of-the-art baselines with respect to F1-score across datasets and programming languages. In particular, MAS-SZZ achieves substantial gains, improving F1-score by up to 65.22% over the best-performing baseline. Contributions. This paper makes the following contributions: • We design and implement MAS-SZZ , a multi-agent system-based SZZ algorithm for VIC identification. • We propose several optimizations, including evidence-grounded root cause analysis, intent-driven anchor statement selection, and autonomous repository exploration, to alleviate the limitations of existing SZZ algorithms. • We evaluate MAS-SZZ against seven existing SZZ algorithms on the two human-annotated datasets. The extensive experiment results indicate that MAS-SZZ outperforms the state-of-the-art baselines across programming languages. Paper Structure. Section 2 introduces the background on SZZ algorithms and LLM-based agents. Section 3 describes the details about our approach. Section 4 presents the experimental setup, followed by the evaluation results. Section 5 concludes the paper. 2. Background and Related Work SZZ Algorithms. The initial SZZ algorithm, B-SZZ ( 13 ) , is based on the premise that “ a bug-inducing commit adds lines that are later removed by a fix” . It traces deleted or modified lines back to their most recent commit. Over the past two decades, numerous variants ( 7 ; 4 ; 3 ; 9 ; 16 ) have been proposed to improve performance. AG-SZZ ( 7 ) filters out cosmetic changes such as whitespace and comments based on annotation graphs to significantly reducing false positives, while MA-SZZ ( 3 ) extends this by excluding meta-changes (e.g., branch merges and property changes) that do not alter program behavior. Recently, LLM4SZZ ( 16 ) further leverages LLMs to assess blame candidates with expanded context, achieving state-of-the-art results across multiple datasets. As an early attempt to extend SZZ algorithms for VIC identification, Nguyen et al. ( 10 ) applied B-SZZ, serving as a foundational starting point for subsequent research. Bao et al. ( 1 ) later proposed V-SZZ, an improved approach that labels the earliest commit as the VIC by computing line similarity. LLM-Based Agents. LLM agents, hereinafter also referred to as agents for short, involve LLM applications that can execute complex tasks through the use of an architecture that combines LLMs with key modules like planning and memory. When building agents, an LLM serves as the main controller that controls a flow of operations needed to complete a task or user request. In recent years, a number of agent-driven solutions have been proposed to solve real-world software problems such as vulnerability detection ( 23 ) and GitHub issue resolution ( 17 ) . For example, Claude Code 1 1 1 https://github.com/anthropics/claude-code is a widely adopted proprietary coding agent, while SWE-agent ( 19 ) bridges LLMs and terminal environments by structuring tool interactions. 3. Methodology Figure 1. Overview of MAS-SZZ . 3.1. Overview As shown