RugKeeper: A Multi-Agent LLM Framework for Rug Pull Token Detection

Abstract

The growth of decentralized finance (DeFi) has been accompanied by an increase in rug pull scams, in which developers misappropriate investors’ funds, rendering the associated tokens worthless. Existing detection methods struggle to capture dynamic on-chain information and provide interpretable risk assessments. This paper presents RugKeeper, a multi-agent framework leveraging large language models for rug pull detection. RugKeeper constructs comprehensive token contexts via a two-step question-driven process and performs multi-path collaborative reasoning, with a Judger Agent validating results to reduce model hallucinations. Evaluations on historical datasets demonstrate that RugKeeper outperforms state-of-the-art methods, achieving 93.55% accuracy, 95.92% F1-score and robust generalization across model backbones. In a real-world sampled dataset from the BNB Chain, 638 previously undetected rug pull tokens were identified. These results highlight the effectiveness of RugKeeper in enhancing DeFi security and supporting risk mitigation.