
LLM-Powered Differential Testing Uncovers 98 Bugs in Ethereum Infrastructure
The video presents research on fortifying Ethereum infrastructure through large language model (LLM)-empowered differential testing, led by Yuho (PolyU, Member Group) and JV Ma (Beihang University). It highlights Ethereum’s $260B market cap, its reliance on smart contracts, and over 300 security incidents in the past year causing nearly $2B in losses. The team used differential testing across 30 nodes (6 consensus-layer and 5 execution-layer clients) to detect inconsistencies in EVM behavior and API responses, uncovering 98 bugs (26 EVM, 72 API), with 85% confirmed or fixed. Case studies include a shallow copy bug in Geth causing chain forks, a Prism bug reducing network participation to 75%, and spec-level errors in the Beacon API. The approach leverages LLMs to generate valid bytecode seeds, extract interprocedural control flow graphs (ICFG), and automate root cause analysis. All 11 mainstream Ethereum clients were affected by the findings, emphasizing the need to prioritize client security alongside smart contract security.