text-to-speech

Pass

Audited by Gen Agent Trust Hub on Jun 18, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill invokes local command-line utilities to perform its core functions.
  • pipeline.py and chunk_and_rewrite.py execute the claude CLI to rewrite text sections via an LLM.
  • kokoro_round5.py and inject_chapters.py call ffmpeg for converting raw audio to MP3 and managing ID3 chapter metadata.
  • inject_chapters.py contains a self-test function that generates temporary audio files using ffmpeg.
  • [PROMPT_INJECTION]: The skill has a surface for indirect prompt injection because it interpolates user-provided text into LLM prompts for narrative rewriting.
  • Ingestion points: The input markdown file is read in pipeline.py and chunk_and_rewrite.py.
  • Boundary markers: The system prompts use ---SOURCE--- delimiters to separate instructions from input text, but do not explicitly instruct the model to ignore any commands found within the source.
  • Capability inventory: The skill can execute local commands (claude, ffmpeg) and write to the local file system.
  • Sanitization: No specific sanitization or filtering is applied to the input markdown before it is sent to the LLM.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 18, 2026, 03:52 AM
Security Audit — agent-trust-hub — text-to-speech