agency-docs-updater

Warn

Audited by Gen Agent Trust Hub on Mar 4, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill orchestrates a sequence of complex tasks by executing local scripts and system binaries (e.g., sync.sh, download_video.py, process_video.py, update_meeting_doc.py, git, gh). These operations are directed to run without human confirmation.
  • [PROMPT_INJECTION]: The script update_meeting_doc.py performs a translation task by interpolating the contents of a meeting summary directly into a command-line prompt for the claude CLI. This creates a surface for indirect prompt injection if the summary (derived from a Fathom transcript) contains malicious instructions.
  • Ingestion points: summary_file read in update_meeting_doc.py and interpolated into a CLI prompt.
  • Boundary markers: Absent; the content is placed directly into the prompt string.
  • Capability inventory: File writing, Git commits, and GitHub API interactions via gh CLI.
  • Sanitization: None; input content is used raw in the command construction.
  • [EXTERNAL_DOWNLOADS]: The skill invokes a Python script to download video files from external Fathom URLs.
  • [PROMPT_INJECTION]: The skill instructions explicitly bypass standard safety checks by directing the agent to 'execute ALL steps... automatically in sequence' and 'Do not stop to ask for confirmation', which removes human-in-the-loop safeguards for a complex multi-stage pipeline.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 4, 2026, 09:29 PM