clean-code-inspector

Fail

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: HIGHPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • Indirect Prompt Injection (HIGH): The skill possesses a significant attack surface by ingesting external content (code diffs) and passing it to sub-agents.
  • Ingestion points: The sub-agent task prompt interpolates file content fetched via git diff -- [파일].
  • Boundary markers: Absent. The sub-agent prompt lacks delimiters (e.g., XML tags or triple quotes) to separate instructions from the code being analyzed.
  • Capability inventory: The skill can spawn sub-agents ('general-purpose'/'sonnet'), execute shell commands (git), and write files to the project root (clean-code-inspect-result.md).
  • Sanitization: Absent. There is no filtering or instruction to the sub-agent to ignore natural language commands embedded in code comments.
  • Command Execution (LOW): The skill uses git status and git diff to inspect the local environment. While these are necessary for the skill's functionality, they interact directly with the host filesystem and repository state.
  • Metadata Deception (LOW): The skill uses strong imperative language ('CRITICAL: You must evaluate...') to enforce its internal framework. While not a safety bypass, this style of instruction is often used in prompt injection patterns.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 16, 2026, 08:48 AM