bailian-multimodal-skills
Pass
Audited by Gen Agent Trust Hub on Apr 16, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill configuration downloads and executes the
uvtool installation script fromhttps://astral.sh/uv/install.sh. Astral is the well-known vendor for this tool, and the process follows standard installation practices for this package manager. - [DATA_EXFILTRATION]: The skill transmits user-provided text prompts and local media file contents (images, audio, and video) to the Aliyun DashScope API endpoint at
dashscope.aliyuncs.com. This data transmission is the core intended functionality for providing multimodal AI features. - [PROMPT_INJECTION]: The skill architecture presents a surface for indirect prompt injection.
- Ingestion points: Processes untrusted external data from user-supplied prompts, remote URLs, and local media files within
run_multimodal.py. - Boundary markers: External inputs are encapsulated in JSON structures for API transmission, but the skill does not use specific delimiters or instructions to protect the model from instructions embedded within processed media or strings.
- Capability inventory: The skill's scripts have the ability to read arbitrary local files provided via command-line arguments and can perform network requests to download generated assets.
- Sanitization: Input data is processed directly as provided by the user or external sources without sanitization or validation against adversarial patterns.
Audit Metadata