Skills
Navigate
skills/qwenlm/qwen-mm-plugins/qwen-mm-plugins-api/Gen Agent Trust Hub

qwen-mm-plugins-api

Pass

Audited by Gen Agent Trust Hub on Oct 30, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted media files (images, videos, audio) from local paths or URLs, which could contain adversarial content intended to influence agent output or behavior.
  • Ingestion points: Multiple tools in SKILL.md such as vision_chat, perceive_media, and omni_asr ingest external media files.
  • Boundary markers: No explicit instruction delimiters or 'ignore embedded instructions' warnings are provided for the model input.
  • Capability inventory: The skill utilizes tools that can read files, extract text, and perform spatial/temporal analysis on media.
  • Sanitization: No explicit sanitization or validation of the media content (beyond format validation) is described.
  • [COMMAND_EXECUTION]: The transcribe_audio tool requires the ffmpeg system utility for extracting and chunking audio from video files.
  • [EXTERNAL_DOWNLOADS]: The reference server implementation references/launch_sam3_server.py specifies dependencies on several well-known technology packages including fastapi, uvicorn, pillow, numpy, and torch.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 30, 2026, 01:45 PM