ANVIL reframes software vulnerability detection as an anomaly detection problem, eliminating the need for labeled training data while achieving superior results to direct LLM-based approaches.

• Uses unsupervised learning to identify code vulnerabilities
• Achieves higher accuracy than direct LLM vulnerability predictions
• Reduces false positives compared to traditional methods
• Provides a practical solution for the labeled data shortage in security applications

This research matters for security teams who struggle with insufficient vulnerability training data and the high costs of false positives. ANVIL's approach offers a more scalable and reliable method for identifying potential security vulnerabilities in software code.

ANVIL: Anomaly-based Vulnerability Identification without Labelled Training Data