LLMs use "safety" specific neuron layers to identify vulnerabilities in code (opens in new tab)

decoderesearch/circuit-tracer: https://github.com/decoderesearch/circuit-tracer

ScholarlyArticle: "Dissecting the Black Box: Circuit-Level Analysis of LLM Vulnerability Detection" (2026-05) https://arxiv.org/abs/2605.29901v1