Argus: Multi-Agent Framework Detects Sensitive Information Leakage with 94.86% Accuracy Via Hierarchical Relationships

Sensitive information leaking from code repositories presents a significant and ongoing security threat, and current detection methods frequently generate unacceptable numbers of false alarms, overwhelming developers. To address this challenge, Bin Wang from Guangdong Provincial Key Laboratory, Shenzhen Graduate School, Peking University, alongside Hui Li and Liyang Zhang from University of Electronic Science and Technology of China, and colleagues, developed Argus, a novel multi-agent framework that dramatically improves the accuracy of sensitive data leak detection. Argus employs a sophisticated three-tier system, analysing code content, file context, and relationships within a project, to minimise false positives and achieve a high level of precision. The team demonstrates Argus’s effectiveness with newly created benchmarks and analysis of real-world repositories, achieving up to 94. 86% accuracy in leak detection, alongside a remarkably low operational cost, representing a substantial advance in automated security practices.

Reducing False Positives in Secret Detection

Scientists are tackling the growing problem of sensitive information, such as API keys and passwords, being accidentally exposed in public code repositories like GitHub. Existing tools often generate numerous false positives, overwhelming developers. Researchers developed Argus, a novel framework employing multiple intelligent agents to detect leaks with greater accuracy and reduce these false alarms. This multi-agent system allows each agent to specialize in a particular detection task, combining their strengths for a more comprehensive analysis. The research positions Argus as a significant step forward in sensitive information leak detection, offering a more robust and accurate solution than existing tools. By combining the power of multiple agents and contextual analysis, Argus addresses a critical security challenge and paves the way for more intelligent and automated software development practices.

Multi-Agent System Detects Code Information Leakage

Scientists developed Argus, a collaborative framework of multiple agents designed to detect sensitive information leaks in code repositories. Recognizing the limitations of traditional methods, the team engineered a three-tier detection mechanism that integrates key content analysis, file context understanding, and project reference relationship mapping to significantly reduce false positives and enhance overall accuracy. This innovative system leverages large language models within a multi-agent architecture, capitalizing on their superior text comprehension capabilities. Rigorous evaluation using newly developed benchmarks demonstrates Argus achieves up to 94.

86% accuracy in leak detection, with a precision of 96. 36%, recall of 94. 64%, and an F1 score of 0. 955. Furthermore, analysis of 97 real repositories incurred a total cost of only $2. 21, demonstrating the system’s efficiency and scalability. The team has made all code and datasets publicly available to facilitate further research and application in this critical area of software security.

Argus Detects Code Leaks with High Accuracy

Scientists developed Argus, a multi-agent framework designed to detect sensitive information leaks in code repositories, achieving a peak accuracy of 94. 86%. The research addresses a critical security challenge, as sensitive information leakage incidents on platforms like GitHub have increased significantly. Argus employs a three-tier detection mechanism, integrating analysis of key content, file context, and project reference relationships to minimize false positives and enhance overall detection precision. This approach overcomes limitations of traditional rule-based tools, which often suffer from high false positive rates, and machine learning methods that lack a comprehensive understanding of code semantics.

Experiments using a newly developed benchmark dataset demonstrate Argus achieves a precision of 96. 36%, a recall of 94. 64%, and an F1 score of 0. 955. The team measured the cost of analyzing 97 real repositories at only $2.

21, demonstrating practical feasibility. By combining multiple AI agents, each focusing on a specific detection task, Argus compensates for the limitations of relying on a single large language model. This collaborative approach delivers stable and precise detection outcomes, significantly outperforming existing methods in identifying sensitive data within complex codebases.

Argus Detects Leaked Secrets With High Accuracy

Argus represents a significant advancement in the detection of sensitive information leakage from code repositories. Researchers developed a multi-agent collaborative framework that substantially reduces false positives, a persistent challenge for traditional detection methods relying on simple pattern matching. The system employs a three-tiered detection mechanism, moving beyond basic key identification to incorporate file context and project-level relationships, thereby improving accuracy and minimizing the burden on developers. Experimental results demonstrate Argus achieves high performance, with an accuracy of 94.

86%, precision of 96. 36%, recall of 94. 64%, and an F1 score of 0. 955 when tested on real-world repositories. Importantly, the system’s analysis of 97 repositories incurred a minimal cost of only $2.

1. The team validated Argus’s capabilities through the creation of new benchmarks specifically designed to assess both leak detection and false-positive filtering. The code and datasets used in this research are publicly available, facilitating further investigation and application of this innovative approach to sensitive information detection.