The core downside: Context vs. guidelines
Conventional SAST instruments, as we all know, are rule-bound; they examine code, bytecode, or binaries for patterns that match recognized safety flaws. Whereas efficient, they usually fail in the case of contextual understanding, lacking vulnerabilities in complicated logical flaws, multi-file dependencies, or hard-to-track code paths. This hole is why their precision charges and the share of true vulnerabilities amongst all reported findings stay low. In our empirical research, the extensively used SAST software, Semgrep, reported a precision of simply 35.7%.
Our LLM-SAST mashup is designed to bridge this hole. LLMs, pre-trained on large code datasets, possess sample recognition capabilities for code habits and a data of dependencies that deterministic guidelines lack. This enables them to cause concerning the code’s habits within the context of the encompassing code, related information, and all the code base.
A two-stage pipeline for clever triage
Our framework operates as a two-stage pipeline, leveraging a SAST core (in our case, Semgrep) to establish potential dangers after which feeding that data into an LLM-powered layer for clever evaluation and validation.
