Evaluating Open-Weight LLMs for Generating Structured Threat Information for Autonomous Vehicle Vulnerabilities

This paper introduces CAV-STIXGen, a dataset that maps connected and autonomous vehicle vulnerability descriptions to STIX Domain Objects, STIX Relationship Objects, CWE labels, and MITRE ATT&CK techniques. It evaluates 11 open-weight LLMs ranging from 4B to 120B parameters under different prompting strategies and temperatures. Single-model systems reach F1 scores of 0.94 for SDO generation, 0.63 for SRO generation, and 0.99 for CWE mapping, while complete ATT&CK mapping remains difficult. In a multi-agent setup, Gemma-4-31B and Codestral-22B achieve SDO and SRO F1 scores of 0.91 and 0.43, respectively.