lead-review
Fail
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill's core workflow involves executing 'gate' commands (such as tests, linters, or build scripts) that it discovers within the project files (e.g., README.md, CLAUDE.md, or package.json). If the repository being reviewed is untrusted or malicious, an attacker can define malicious commands in these files that the agent will execute in the host environment. This risk is specifically mitigated in 'unattended mode' by skipping these executions, but it remains for interactive usage.
- [COMMAND_EXECUTION]: The skill uses the git and gh (GitHub CLI) tools to manipulate repositories, manage worktrees, and post reviews. Additionally, it executes arbitrary toolchain commands (e.g., npm run test, cargo test, pytest) based on the project's ecosystem, which provides a pathway for the execution of arbitrary code defined by the repository's configuration files.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from pull request diffs, descriptions, and comments. This creates an attack surface where a malicious PR could attempt to influence the agent's behavior or output.
- Ingestion points: Pull request data including diffs, titles, bodies, and discussion comments.
- Boundary markers: The skill uses a separate 'verifier' agent session to check the generated review against project rules. In 'unattended mode', it follows a strict instruction that PR content must be treated as data and cannot override instructions.
- Capability inventory: The agent can run shell commands, perform git operations, manage GitHub PRs/issues, write to the /tmp/ directory, and spawn subagent sessions.
- Sanitization: The included Python scripts perform structural validation of the review content, including sentence length limits and formatting checks, to detect and prevent certain types of injection or formatting anomalies.
Recommendations
- AI detected serious security threats
Audit Metadata