OWASP 2025 Global AppSec EU

False positives are one of the biggest pain points in running a Static Application Security Testing (SAST) program. While SAST tools are valuable for identifying security issues in a codebase—flagging critical vulnerabilities like Remote Code Execution and SQL Injection—they often generate significant noise due to their lack of contextual awareness. SAST testing is relatively easy to set up, requires no accounts or credentials, and can uncover issues in multi-step processes that would be difficult to detect with dynamic security testing. However, the high volume of false positives leads to alert fatigue and demands considerable effort to triage, making it challenging to identify the relatively small number of true vulnerabilities.

This research addresses that challenge by combining Program Analysis with Large Language Models (LLMs) to simulate the manual triage process for SAST findings. Our approach leverages a carefully designed LLM agent that enhances context around vulnerable code, identifies conditions that make exploitation infeasible, and determines whether a clear execution path exists from a user-controlled input to the vulnerable line flagged by SAST.

We will demonstrate this novel approach in action, showcasing how it can be integrated with any SAST tooling to streamline triage. By reducing false positives and prioritizing actionable findings, this method allows security engineers and developers to focus on the vulnerabilities that truly matter.