AI Now Finds Flaws in Your Crypto Wallet (and Hackers Can Too)

Researchers used large language models to automatically inject vulnerabilities into Solidity smart contracts, aiming to help assess vulnerability detection tools. Applied to real-world contracts, the LLMs generated nearly 1,000 candidate variants; after validation, 32 confirmed vulnerable contracts spanning 25 vulnerability types survived—a 16.58% survival rate. Survivors were concentrated in structurally simpler contracts and vulnerability types with localized syntactic patterns. The authors also report practical challenges, including LLM non-determinism and difficulty preserving contract semantics. Using the validated contracts to assess three static analyzers revealed complementary but incomplete coverage. Overall, the article finds LLM-based vulnerability injection is feasible, while exposing key limitations in scalability and diversity.