qwen-mm-plugins-video-memory

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). The skill’s required runtime workflow reads and ingests the pre-built long-video memory files (graph_memory.json/subgraphs/embeddings) from <video_path>.memory/ and runs query tools (search_nodes/search_ocr_text/search_asr_text/get_subgraph) over that contents, so outsider-authored text can be injected if an outsider can supply or poison those stored memory artifacts.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The code calls DashScope APIs at runtime (DEFAULT_BASE_URL "https://dashscope.aliyuncs.com/compatible-mode/v1" and the native host "https://dashscope.aliyuncs.com") which generate the model JSON used to build the graph memory, and the build script may fetch and install Python packages from "https://mirrors.aliyun.com/pypi/simple/" during runtime, so external content is fetched and directly influences outputs and executes remote code via installs.