Original title: LLMSec-AV: A Vulnerability Taxonomy and LLM-Driven Software Weakness Discovery Framework for Autonomous Vehicles
Md. Wasiul Haque, Sagar Dasgupta, Mizanur Rahman · September 2026
Self-driving cars run on millions of lines of code. If that code fails, it can alter how the car moves. Normal security tools struggle to spot these flaws because they do not understand vehicle safety. To fix this, researchers created a new system called LLMSec-AV. They made an automated-vehicle vulnerability taxonomy, which is a list of 18 specific types of weaknesses drawn from past car security reports. Then, they had large language models review real vehicle software. The models analyzed code functions under several setups, using the weakness list, past reports, and multi-step reviews. The models found up to 76% of 46 known weaknesses in the software, easily beating traditional tools. Standard tools missed almost all of them. Interestingly, the weakness list did not help the model spot more flaws. Instead, it helped categorize over 80% of the findings into specific weakness classes, making them much easier for engineers to interpret.
Imagine you're a detective trying to find hidden traps in a huge mansion. You have a checklist of common trap types, but the mansion has thousands of rooms. You could search every room, but that takes forever. Instead, you use a smart assistant who reads the blueprints and points out the most likely spots. In this paper, the LLM is that assistant, and the taxonomy is the checklist. It doesn't find every trap, but it finds far more than a random search, and it labels each trap with its type, making it easier to fix.
Autonomous vehicles run on millions of lines of safety-critical code. If a flaw hides in the part that controls motion, lives are at stake. This research shows that large language models, guided by vehicle-specific security knowledge, can catch weaknesses that standard code analyzers miss, potentially making self-driving cars safer.
The abstract does not report precision or false-positive rates, so we don't know how many alerts were false alarms. It also does not compare against other LLM-based approaches or test on multiple AV systems. The taxonomy did not improve recall, suggesting the LLM's prior knowledge was sufficient. Six of the 18 classes could not be represented as static-analysis rules, limiting direct integration.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
LLMSec-AV: A Vulnerability Taxonomy and LLM-Driven Software Weakness Discovery Framework for Autonomous Vehicles — Md. Wasiul Haque, Sagar Dasgupta, Mizanur Rahman
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.