ATLAS: Discovering Agent Strategies through LLM-Guided Abstraction and Automata Learning
Source
arxiv.org
Author
Ignacio D. Lopez-Miguel, Andreas Happe, J\"urgen Cito, Ezio Bartocci, Bettina K\"onighofer, Martin Tappler
Date
Context
ATLAS turns records of an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition →'s actions into a simpler map of its observed behavior. With a LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → guiding the grouping, it combines detailed interactions into higher-level states and learns transitions between them, producing a finite-state model that people can inspect for recurring strategies, successful paths and failure loops.
The researchers demonstrate the approach on a penetration-testing agent, including a case study involving 12 vulnerable machines. Its value is making patterns across action sequences easier to examine than raw logs alone. The resulting map describes observed behavior; it does not reveal hidden thoughts or establish how the agent will behave in every new situation.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
articleAgentic Security: A Systematization of Tools, Failure Modes, and Design Laws for LLM-Driven Penetration TestingIsrat Moyeen Noumi, Tarannum Ahmed Nowshin, Md. Mehedi Hasan Nipu, Mohammad Sakib Mahmood, Md. Jakir Hossain, M. F. Mridha
articleAgentic Data Engineering for LLMs: System Design and a Controlled Empirical StudyRuomeng Ding, Qianli Shen, Zhaoyang Han, Meixin Chen, Daoyuan Chen, Yaliang Li
articleAgent Stack for ML Training Lifecycle: Generation, Validation, and OptimizationJaideep Ray, Daniel Tang, Zejun Lin, Haoyue Tang, Lijuan Zhang, Tao Huang, Tommy Li, Aditya Jain, Ankit Goyal, Harsh Jain, Zhuo Zhi, Youmin Han, Santosh Jha, Shuying Liang, Yujie Ai, Jonathan Hung, Ethan Lyu, Abi Liu, Siddharth Singh, Feng Wen, Mahesh Ravindran, Souvik Ray, Yitong Zhou, Animesh Singh