toutiao-publisher

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The publisher.py script uses subprocess.run to interact with system utilities such as xclip, wl-copy, xdotool (Linux), osascript, swift (macOS), and powershell.exe (Windows) to handle clipboard operations for pasting images into the browser editor.
  • [EXTERNAL_DOWNLOADS]: The setup_environment.py script automates the installation of dependencies from official registries (PyPI) and downloads browser binaries (Google Chrome) required for the patchright library.
  • [DYNAMIC_EXECUTION]: The publisher.py script dynamically generates and executes platform-specific code (Swift on macOS and PowerShell on Windows) to perform clipboard operations that are not natively supported by standard Python libraries.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external Markdown files for publishing, which serves as an ingestion point for untrusted content. While it possesses capabilities such as network access (via browser automation) and command execution (via clipboard tools), the risk is mitigated by its intended primary purpose and user-initiated execution flow.
  • Ingestion points: Reads content from local Markdown files via --content-file in publisher.py.
  • Boundary markers: The content is processed as article data for conversion to HTML; no specific LLM boundary markers are employed to distinguish instructions within the text.
  • Capability inventory: The skill can perform network operations via browser automation, write/read local files in the skill data directory, and execute system commands for clipboard management.
  • Sanitization: Content is converted to HTML using a standard Markdown library, and the agent uses a human-review mode (manual) by default to verify output before publication.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 07:59 AM