critical-thinking
Pass
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes an internal Python script,
scripts/coverage.py, to verify that its logical analysis captures every relevant line of the input document. The script is executed locally via theuvtool and uses secure practices such asyaml.safe_load()for data handling. - [INDIRECT_PROMPT_INJECTION]: As a document auditing tool, the skill naturally ingests untrusted text from user-provided files. This constitutes a potential attack surface for indirect prompt injection. The skill mitigates this risk by converting input into a rigid YAML-based logical structure, treating user content as data to be analyzed rather than instructions to be followed.
- Ingestion points: External files specified by the user for auditing (e.g.,
<対象ファイル>). - Boundary markers: The workflow mandates mapping every piece of input text to a structured YAML schema (facts, inferences, etc.) and specific line numbers.
- Capability inventory: File system read/write operations for generating reports and execution of the local
coverage.pyscript. - Sanitization: The analytical process requires the agent to decompose the text into logical components, which prevents embedded instructions from being interpreted as agent commands.
Audit Metadata