ATLAS Recovers Finite-State Strategy Models From Agent Traces, Applied to an LLM Penetration-Testing Agent

arXiv 2608.14352·medium signalXBlueskyLinkedInCopy link

ATLAS combines trace abstraction with automata learning to infer interpretable finite-state models of what strategy an LLM agent is actually following, going past the usual task-success-plus-raw-traces evaluation to expose recurring behaviors, decision points, successful completion paths, and failure loops. Demonstrated on an LLM penetration-testing agent across 12 vulnerable machines, the learned models surface exploitation strategies that are not visible in raw execution traces. The authors also show symbolic model-based knowledge transfer from frontier models to compact ones — the behavioral model becomes a portable artifact, not just an audit report.