editor-apprentice
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted audio and video files. It transcribes speech and uses the resulting text to influence the editing process. Rule 6 in SKILL.md explicitly directs the agent to follow 'spoken directions' found within the recordings, creating a vector for indirect prompt injection if an attacker embeds malicious instructions in a video's audio track. Mandatory Evidence Chain: (1) Ingestion points: transcription of media by transcribe.py into .islands.json files; (2) Boundary markers: None; (3) Capability inventory: Subprocess calls for system commands, file system writes, and ExtendScript execution; (4) Sanitization: None.
- [COMMAND_EXECUTION]: The skill utilizes the subprocess module across multiple scripts to execute external commands. While these are legitimate for the skill's purpose (e.g., ffmpeg for media processing, AfterFX.exe for application control), the availability of these commands to an agent represents a significant capability.
- [EXTERNAL_DOWNLOADS]: The setup_check.py script facilitates the installation of third-party Python packages (numpy, faster-whisper, opencv-python-headless) from PyPI. Additionally, transcribe.py downloads pre-trained transcription models from Hugging Face.
- [DYNAMIC_EXECUTION]: The skill generates and executes code at runtime. Specifically, ae_run.py creates ExtendScript (.jsx) wrappers to drive After Effects. Furthermore, resolve_bridge.py dynamically loads the DaVinci Resolve scripting API from the local installation directory.
Audit Metadata