code-review-multi-axis
Pass
Audited by Gen Agent Trust Hub on Oct 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill's core functionality involves analyzing untrusted content from various sources which could contain malicious instructions.
- Ingestion points: The skill retrieves data from PR descriptions, source code diffs, commit messages, and issue tracker content using
gh pr viewandgh issue viewcommands. - Boundary markers: The instructions for the parallel sub-agents defined in
PRE-PR.mddo not explicitly mandate the use of robust delimiters (such as XML tags or unique block markers) to separate the untrusted diff and spec content from the analysis instructions. - Capability inventory: The agent has permissions to read all files in the local repository, create git worktrees, and perform authenticated network requests via the GitHub API.
- Sanitization: The instructions lack specific requirements for sanitizing or escaping the content fetched from external sources before it is processed by the model or used in the review approval loop.
- [COMMAND_EXECUTION]: The skill frequently executes shell commands that incorporate user-provided identifiers and references.
- Pattern: The skill instructs the agent to run commands such as
gh pr view <id>,git diff <ref>, andgit worktree add <path> <sha>. - Risk: User-supplied strings (PR numbers, branch names, or commit SHAs) are used as arguments. While the skill includes validation steps (e.g.,
git rev-parse), there is a potential for command injection if the execution environment or the agent's shell handling does not correctly sanitize these inputs.
Audit Metadata