When Models Know but Can't Linearly Say: Relational Encoding Limits Linear Vulnerability Classification in Code LLMs

Large language models (LLMs) can detect software vulnerabilities, but how do they actually identify vulnerable code? We address this question using mechanistic interpretability; analyzing the internal computations of a neural network to understand its reasoning process.Using Circuit Tracer on Gemma-2-2b, we trace the computational pathways activated when the model classifies 472 C/C++ code samples as vulnerable or safe. Our analysis reveals a surprising finding: the model primarily relies on safety detectors, attention heads that recognize safe coding patterns, rather than directly detecting vulnerability signatures. When these safety detectors fail to activate, the model classifies code as vulnerable. We identify the critical neural components: specific attention heads in early layers (L5, L7) that focus on safety patterns, and Multilayer Perceptron (MLP) neurons in Layer 7 that encode vulnerability-related features. Ablation experiments confirm their causal role; removing Layer 11 drops vulnerability detection accuracy from 100% to 6%, while ablating just 20 neurons in Layer 7 reduces it by 50%.Our findings show that LLM vulnerability detection uses sparse, interpretable circuits (only 16% of model capacity), enabling circuit-level explanations for security predictions and targeted improvements to detection systems.

Researchers and practitioners increasingly apply Large Language Models (LLMs) for automated vulnerability detection. Recent work has shown that LLMs are susceptible to the same cognitive heuristics that bias human judgment. Yet, no work has investigated whether these heuristics affect a model's assessment of code vulnerabilities. In this paper, we present the first systematic exploration of cognitive heuristics in LLM-driven code vulnerability detection. We introduce a controlled framework that holds the code fixed and only varies the surrounding context to trigger three cognitive heuristics: the halo effect through author attribution, the framing effect through task objectives and consequences, and the anchoring effect through prior analysis results. Within this framework, we evaluate eight LLMs across three programming languages. Our results demonstrate that all evaluated models are susceptible to these heuristics. Cross-model average susceptibility is highest for framing at 33.2%, followed by anchoring at 23.5% and halo at 18.4%. Vulnerabilities that require semantic reasoning for detection are more susceptible to cognitive heuristics than those identifiable through pattern matching. Furthermore, models often change their verdict from safe to vulnerable based on the cognitive condition, without accurately identifying the actual vulnerability. To highlight the practical impact, we demonstrate a proof-of-concept black-box cognitive attack that can suppress up to 97% of previously detected vulnerabilities. Together, these findings show that cognitive bias is a consistent and exploitable property of LLM-based vulnerability detection.

LLM-based code generation is now embedded in mission-critical pipelines, but defenses against vulnerable output remain post-hoc -- static analyzers, fine-tuned classifiers, or an LLM judge that screen completed code, ignoring the generating model's own internal state. We test a narrower, directly measurable question: when an LLM reads a piece of C/C++ code as context, do its hidden activations already carry a signal about that code's vulnerability status? We extract last prefill token activations from four LLMs (Granite-4.1-8B, Qwen3.5-9B, Qwen3.6-27B, Gemma-4-12B) across three model families and train MLP probes on these activations. We evaluate them on four function-level C/C++ benchmarks (Devign, Big-Vul, Draper VDISC, PrimeVul). Our probes achieve 41.7% average F1 using 13.4--16.0M-parameter probes -- under 0.2% of base-model size. On Devign, the best probe (Qwen3.5-9B, 68.8% F1) matches the published fine-tuned-classifier SOTA (67.9%) despite reading only a frozen, general-purpose LLM's activations; on the harder, more imbalanced benchmarks (Big-Vul, Draper VDISC, PrimeVul) probes trail SOTA substantially. This is early evidence that a coding LLM's own representation of arbitrary code is informative about that code's vulnerability status, motivating further work toward lightweight, model-native vulnerability screening.