adversarial / code-pr

Adversarial code review evaluates proposed AI-generated diffs by running independent critics for structural issues, test weaknesses, dependency risks, and data leaks. It assesses pull requests against plan success criteria and test outcomes to produce a deterministic ship or hold decision.

Slopsquatting risk in AI code dependencies occurs when AI-generated code introduces unverified or malicious packages. Adversarial review checks that new dependencies map to verifiable entries in their registries, preventing this supply chain attack vector before the pull request ships.

Preventing data leaks in AI-generated code diffs requires targeted exposure surface analysis. The review process runs a dedicated data-leak critic over the proposed changes to identify and block potential sensitive information exposure before approving the pull request.

Pull request validation adjusts review depth dynamically based on a structural risk score. High-risk diffs trigger exhaustive per-test inspection and multiple critic iterations, while lower-risk changes receive a standardized evaluation pass.

Common AI failure modes in generated code changes include structural defects, insufficient test coverage, and insecure data handling. A dedicated critic evaluates the diff specifically for these AI-specific patterns to catch issues that standard linters miss.

You should hold an AI-generated pull request from shipping when the adversarial review identifies any blocker-level findings. The evaluation outputs explicit severities for structural issues, test weaknesses, dependency risks, and data leaks to enforce evidence-driven readiness.