notebooklm

Pass

Audited by Gen Agent Trust Hub on Sep 12, 2026

Risk Level: SAFECOMMAND_EXECUTIONCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses a central wrapper script (scripts/run.py) to manage its internal environment and route commands to specific automation scripts.
  • The run.py script employs os.execv() to switch the process context to a local virtual environment.
  • Security controls include a strict allowed-list of supported scripts (ALLOWED_SCRIPTS) and a regular expression check for script filenames.
  • Path resolution checks are performed to ensure script execution cannot escape the designated scripts/ directory.
  • [CREDENTIALS_SAFE]: The skill manages Google authentication data with robust local security measures.
  • Authentication cookies and session state are stored in a dedicated user data directory (~/.local/share/agentic-awesome-skills/notebooklm).
  • The skill sets a global os.umask(0o077) to ensure all files created by the browser or the skill are private by default.
  • Explicit chmod calls restrict sensitive files (state.json, auth_info.json) to 0600 permissions and directories to 0700 permissions.
  • [INDIRECT_PROMPT_INJECTION]: The skill acknowledges the risk of processing untrusted content from the NotebookLM web interface.
  • Ingestion points: External data enters the agent context via responses scraped from notebooklm.google.com in scripts/ask_question.py.
  • Boundary markers: The skill uses explicit delimiters (--- BEGIN UNTRUSTED NOTEBOOKLM CONTENT ---) to separate external data from system instructions.
  • Capability inventory: The skill has the ability to write to private local storage and execute internal scripts via the run.py wrapper.
  • Sanitization: External content is serialized as JSON strings and written to private temporary files to prevent it from polluting terminal logs or being accidentally interpreted as commands.
  • Human-in-the-loop: Instructions in SKILL.md require the agent to pause and ask for user confirmation before using metadata (names, descriptions, topics) discovered from a notebook.
  • [EXTERNAL_DOWNLOADS]: The skill automates the installation of necessary browser binaries during its first-run setup.
  • It uses the official patchright CLI to install Google Chrome and Chromium binaries, which is a standard procedure for browser automation tools.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 12, 2026, 09:54 AM
Security Audit — agent-trust-hub — notebooklm