Automated Vulnerability Injection in Smart Contracts Using Large Language Models

Researchers used language models to deliberately introduce security flaws into smart contracts, creating test cases for vulnerability detectors. The useful output was a collection of confirmed defects: each candidate had to pass checks for compilation, execution, preserved business logic and the intended vulnerability. The study reports 32 validated contracts covering 25 vulnerability types after generating nearly 1,000 candidate variants and removing duplicates.

The surviving examples favored simpler contracts and flaws requiring localized code changes. The result shows that AI can help build security-testing datasets, while the validation process determines which generated examples are credible enough to test a scanner.