muapi-director
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external inputs such as user-provided topics and audio transcripts (generated via Whisper ASR) and interpolates them into prompt templates for image and video generation. There is a lack of boundary markers or sanitization in the prompt construction, which allows external content to potentially influence the behavior of the downstream generation models. For example, the
narrationandtopicfields frombeats.jsonare used directly in prompts withinaroll_clips.pyandclips.py. - [COMMAND_EXECUTION]: The orchestration scripts rely heavily on
subprocess.runto execute system binaries, specificallyffmpeg,ffprobe, andcurl. These are used for video assembly, media metadata probing, and file transfers between the local system and remote services. Evidence includes numeroussubprocess.runcalls found inassemble.py,aroll_assemble.py,asr_beats.py, andmuapi_client.py. - [EXTERNAL_DOWNLOADS]: The skill communicates with well-known service endpoints including
api.muapi.aiandapi.openai.comto perform AI generation and transcription tasks. It usescurlto upload local media assets to these services and download the resulting media files.
Audit Metadata