pocketpolyglot-bookmaker

Pass

Audited by Gen Agent Trust Hub on Oct 19, 2026

Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DATA_EXFLLTRATION]: The skill contains a hardcoded absolute file path /home/lachlan/.nutstore/logs/client.log. This identifies a specific system user and exposes the local directory structure of the host machine.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external content from the sources/ directory, including EPUB, PDF, and Markdown files.
  • Ingestion points: Source book files in EPUB, PDF, and Markdown formats provided in the sources/<book> directory (SKILL.md).
  • Boundary markers: The instructions provide guidelines to filter out "HTML, Wikisource boilerplate, PDF page headers, archive labels, copyright/public-domain notices, or OCR noise" from entering the text chunks, which provides some separation but does not fully isolate the agent from embedded instructions.
  • Capability inventory: The skill possesses the ability to read and write files (JSON and PDF), execute shell commands (find, tmux), run repository-specific Python scripts, and invoke the XeLaTeX compiler.
  • Sanitization: The skill mentions filename sanitization for Nutstore compatibility and instructs the agent to remove boilerplate metadata from source text.
  • [COMMAND_EXECUTION]: The skill uses shell commands and local repository scripts to manage book generation and monitoring.
  • Evidence: Usage of find, tmux capture-pane, and execution of scripts like ./studio/pocketpolyglot and scripts/books/nutstore_paths.py. These are documented as core components of the bookmaking workflow.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 19, 2026, 06:25 PM