Initial Attempt at Automated Vulnerability Fixing of Obfuscated Smart Contracts Using a Large Language Model

Abstract

Countermeasures against vulnerabilities are essential for the security of smart contracts. Some recent studies have focused on automated vulnerability repair using large language models (LLMs). However, the ability of LLMs to fix unknown contracts is still unclear. To address this problem, we applied Code Llama LLM to a real-world smart contract with a vulnerability and its obfuscated version, assuming that the obfuscated version could be considered an unknown contract. We obtained three fixes generated by the LLM for each contract. Of these three, one fix was successful for each contract. While some fixes removed the vulnerability, they also altered the functionality of the contract. No clear difference was observed between the original and obfuscated contracts in terms of the LLM's repairing capability.