Can AI models perform smart contract audits?

AI models, such as large language models (LLMs), have made significant strides and can be useful in auditing smart contracts. However, they're not ready to fully replace manual audits. While they can provide a brief overview of how a contract works and help in detecting some simple bugs, they face several issues like the constraints of context window, the inability to detect new vulnerabilities, finding only simple bugs, a tendency to "hallucinate," and a lack of transparency.

What are the issues with using AI for smart contract audits?

AI models face several issues in auditing smart contracts. They have a constraint of a context window that limits the amount of code that can be analyzed at once. They are trained on existing data and known vulnerabilities, so new vulnerabilities might be missed. AI models also tend to find only simple bugs and may "hallucinate" or provide inaccurate information. There's also a lack of transparency where it's not always clear if the model has processed the entire codebase or just a part of it.

What are the main use cases for AI in smart contract audits?

Despite their limitations, AI models can still be very useful in smart contract audits. They can provide a brief overview of how a contract works, which can help understand the contract faster. AI models can also point out areas of code that could potentially be vulnerable and find simple bugs, acting as an extra check to minimize the risk of human error. Besides, these models, which are adept at writing text, can assist in describing manually found issues, especially for non-native speakers. In the future, these models can act as helpful assistants for manual audits, trained to detect specific vulnerabilities.