natural-japanese

Pass

Audited by Gen Agent Trust Hub on Oct 24, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill fetches a large language model from a well-known service for advanced text analysis.
  • Evidence: scripts/semantic.py downloads a ~1GB model from HuggingFace (cl-nagoya/ruri-v3-310m) when the experimental exp mode is activated.
  • Context: This is a documented feature used for "semantic topic flatness" detection.
  • [DYNAMIC_EXECUTION]: The skill utilizes dynamic loading for internal script modularity and model execution.
  • Local Loading: scripts/calibrate.py uses importlib.util to dynamically load and execute the local lint.py script for calibration tasks.
  • Model Execution: scripts/semantic.py uses trust_remote_code=True when initializing the SentenceTransformer model from HuggingFace, which allows loading custom code provided by the model repository.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided text which may contain embedded instructions.
  • Ingestion points: User-supplied Japanese prose for rewriting, scoring, or diagnosis via SKILL.md commands.
  • Boundary markers: The skill emphasizes semantic fidelity and uses reference guides (e.g., references/semantic-fidelity.md) to instruct the agent to prioritize meaning over stylistic changes.
  • Capability inventory: The skill performs file reading and executes several local Python scripts via uv run to analyze text.
  • Sanitization: scripts/textcore.py provides regex-based masking for Markdown structures and HTML comments, though it focuses on excluding non-prose elements from analysis rather than security sanitization.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 24, 2026, 05:28 PM