Study Validates Predicate Granularity in Automated Attack Chain Generation
An empirical investigation published on arXiv (2608.00143) explores how the granularity of predicate representation affects the generation of automated attack chains in cybersecurity. This research contrasts the nine-category Attack Action Linking Model (AALM) utilized by the AURORA system with a simplified five-category model based on Atomic Red Team (ART) execution data. Utilizing a pipeline, a Large Language Model (LLM) translates techniques into symbolic predicates, while the Fast Downward engine executes deterministic reasoning. Findings from a sixteen-technique dataset reveal that the five-category model can match the plan validity, cost, and fidelity of the more detailed nine-category model. These results imply that less granular predicate representations might suffice for effective attack planning, potentially streamlining future automated systems. This study enhances AI planning in cybersecurity by offering empirical insights into the balance between representation granularity and planning efficiency.
Key facts
- Study published on arXiv with identifier 2608.00143
- Compares nine-category AALM against a five-category scheme
- Uses LLM for translation and Fast Downward for reasoning
- Based on Atomic Red Team (ART) execution evidence
- Corpus includes sixteen techniques
- Evaluates plan validity, cost, and fidelity
- Challenges the necessity of the nine-category granularity
- Suggests coarser representations may be sufficient
Entities
Institutions
- arXiv
- Atomic Red Team
- AURORA