What problem does it solve? Code changes that touch trust boundaries β authentication, input parsing, secrets handling, CI workflows β can introduce vulnerabilities that pass functional tests but fail under adversarial conditions. This Skill performs a structured security review of a diff before merge, catching injection, path traversal, secret leakage, privilege escalation, and fail-open guards that reviewers commonly miss. ## Core Features & Use Cases - Adversarial Diff Review: Treats the change as hostile until proven safe, tracing untrusted input from source to sink across shell, SQL, YAML, and LLM instruction contexts. - Trust-Boundary Fan-Out: Splits large changes into per-boundary review passes (auth, input parsing, secrets, CI/exec, file I/O) and dedupes findings. - Verify-Before-Reporting: Requires each finding to state the exact triggering inputs and resulting wrong behavior, defaulting to "refuted" when uncertain to avoid false positives. - Use Case: Before merging a PR that adds a CI workflow running on a self-hosted runner, ask for a security review to check whether untrusted fork code could execute with privileged permissions or whether third-party actions are unpinned. ## Quick Start Ask the AI to security review the current branch diff against main and report ranked findings with file, line, severity, and suggested fixes.
