Adversarial Code Review: Can AI Reviewers Be Fooled by Semantic Perturbations?

AI-powered code review tools are increasingly deployed in production environments to automate software quality assurance. However, the security implications of these systems remain underexplored. This paper investigates whether AI code reviewers can be fooled by adversarially perturbed code, building on prior work that identified a three-tier detectability hierarchy for code perturbations. We evaluate a CodeBERT-based AI reviewer on 550 Python samples containing five perturbation strategies, measuring false approval rates across perturbation types. Our results reveal a critical vulnerability: semantic boundary inversions achieve a 90% false approval rate, meaning AI reviewers approve 9 out of 10 adversarially modified code samples. Even an ensemble system combining AI embeddings with AST features fails to detect 80% of boundary inversions. These findings highlight significant security gaps in AI-based code review systems and underscore the need for specialized detectors targeting semantic perturbations.