iflytek-image-understanding
Pass
Audited by Gen Agent Trust Hub on Oct 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data (images and user-supplied questions) and incorporates them into a prompt for the Spark Vision model, creating a surface for indirect prompt injection.
- Ingestion points: The script
scripts/image_understanding.pyreads image files from the filesystem and accepts strings for questions through command-line arguments. - Boundary markers: The implementation lacks explicit delimiters or specific instructions for the model to ignore potential commands embedded within the image data or question text.
- Capability inventory: The skill has the ability to read local files and perform network operations via WebSockets.
- Sanitization: No sanitization or validation of the input content is performed before it is encoded and transmitted to the API.
- [DATA_EXFILTRATION]: The skill transmits image content to an external service for analysis.
- Evidence: The script
scripts/image_understanding.pysends base64-encoded image data to the WebSocket endpointwss://spark-api.cn-huabei-1.xf-yun.com/v2.1/image. - Context: The endpoint belongs to the vendor's official infrastructure (iFlytek/Xunfei), and the communication is secured using HMAC-SHA256 signatures and TLS/SSL.
Audit Metadata