TrapHunter: New AI framework detects hidden traps in ERC-20 token contracts

TrapHunter addresses the growing problem of "trap tokens"—ERC-20 compliant contracts that follow standard protocols to evade detection while embedding covert logic to defraud users. The researchers first systemize the trap landscape with a novel taxonomy based on the token lifecycle (Generation, Circulation, Persistence, Observation). The framework then combines Abstract Behavior Trees (ABTs) and Augmented Path Graphs (APGs) to normalize intra-procedural syntax and reveal hidden execution paths driven by inter-procedural state dependencies. Crucially, TrapHunter bridges the semantic gap by using LLMs to reason about behavioral intent deviations from reference implementations, followed by fork-based dynamic validation to confirm exploitability.

In experimental evaluation on 269 real-world contracts with three LLMs (DeepSeek, GPT, Gemini), TrapHunter detected all six categories of traps with an average precision of 81.8% and recall of 85.4%, significantly outperforming state-of-the-art tools. The paper has been accepted to ISSTA 2026, the premier software testing conference. This work represents a major advancement in automated smart contract security auditing, particularly for sophisticated scams that rely on "deceptive adherence" rather than obvious code violations. The hybrid approach of static analysis combined with LLM-based intent reasoning could become a new standard for blockchain security tools.