loop-review
Warn
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: MEDIUMREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill executes package installation commands (e.g.,
pnpm install) on the working tree of an untrusted pull request. This triggers the execution of pre-install, post-install, and other lifecycle scripts defined in the pull request'spackage.json, allowing for arbitrary code execution from external sources. - [COMMAND_EXECUTION]: Multiple shell commands are constructed using variables derived from user input or external pull request data (e.g.,
<PR>,<owner>,<repo>, and<number>). Examples includegh pr view <PR>,git worktree add --detach <path>, andgh pr checkout <PR>. If the normalization process or the agent fails to sanitize these inputs, it could lead to command injection vulnerabilities. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon data from external pull requests, creating a surface for indirect prompt injection.
- Ingestion points: Pull request metadata and code diffs are ingested via
gh pr viewandgh pr checkout(documented inSKILL.mdunder the 'Setup' section). - Boundary markers: The instructions do not define clear delimiters or specific instructions to ignore embedded commands within the code or comments being reviewed.
- Capability inventory: The skill possesses significant capabilities, including arbitrary shell command execution during project setup, the ability to load other skills (
code-review-and-quality), and the authority to performgit pushto remote repositories. - Sanitization: There is no evidence of sanitization or validation of the pull request content before it is processed by the agent or committed back to the repository.
- [EXTERNAL_DOWNLOADS]: The skill automates the cloning of external repositories and the installation of their dependencies from public registries using tools like
ghandpnpm. While these are functional requirements, they involve downloading and executing code from external, potentially unvetted sources.
Audit Metadata