Build Integration, Not Fuzzing, Is What Kills LLM Dynamic Analysis on Real Autonomous-Vehicle Stacks

arXiv 2608.13450·medium signalXBlueskyLinkedInCopy link

Compiler-precise static analysis across 185 Autoware packages surfaced 1,375 decision rules, 2,274 validation checks and 482 input-to-safety-output flows, from which 740 reachable weakness sites were sampled and 3,700 LLM-generated artifact sets compiled against the real sanitized build. 80% of first-shot compilation failures came from dependency wiring rather than program logic; a reasoning model compiled 64% of harnesses first try versus 6% for a code-specialized model. Even after compiler-in-the-loop repair, fewer than half of harnesses reached the fuzzer and all 37 observed crashes originated in stubbed code — no candidate weakness was dynamically confirmed within budget.