build-in-public-youtube
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes local system commands to process files, including
textutilfor converting RTF documents andwcfor estimating transcript length. Evidence: 'textutil -convert txt -stdout "input.rtf" > /tmp/transcript.txt' in SKILL.md.\n- [COMMAND_EXECUTION]: The skill invokes a local Python scriptMSBC-Group/tools/transcribe.pyto handle audio files. Evidence: 'python3 MSBC-Group/tools/transcribe.py ' in SKILL.md.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted meeting transcripts or recordings that could contain instructions designed to influence the agent's behavior. Evidence: 'Step 1 — Read the transcript fully' in SKILL.md.\n - Ingestion points: Files read from the user's iCloud TextEdit directory (
~/Library/Mobile Documents/com~apple~TextEdit/Documents/) or local recording folders.\n - Boundary markers: Absent. The instructions do not specify the use of delimiters or 'ignore embedded instructions' warnings when reading transcripts.\n
- Capability inventory: The skill has the capability to execute shell commands, run local Python scripts, and write episode packs to the
/tmp/directory.\n - Sanitization: Absent. There is no mention of escaping or validating the content of transcripts before they are interpolated into the content generation steps.
Audit Metadata