SoK: Understanding (New) Security Issues in AI4Code Use Cases

A SoK paper systematically reviews the security issues of AI4Code in code generation, translation, and vulnerability detection, and compares six frontier models. The paper identifies recurring gaps in the field: benchmarks dominated by Python and toy problems, a lack of standardized security datasets, data leakage in evaluation, and fragile adversarial robustness. A comparative study of six models shows that insecure patterns persist in code generation, vulnerability detection is vulnerable to semantics-preserving attacks, fine-tuning is often misaligned with security objectives, and the security benefits of code translation are mixed. Based on this, the authors propose three paths: embedding secure-by-default practices in code generation, building robust and comprehensive detection benchmarks, and using translation as a route to safer languages, and call for a shift toward security-first AI4Code.