glmv-prompt-gen

Pass

Audited by Gen Agent Trust Hub on Aug 12, 2026

Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill ingests user-provided images and videos which may contain adversarial text or patterns intended to influence the AI's output (Indirect Prompt Injection). \n
  • Ingestion points: Media files and URLs are processed through scripts/prompt_gen.py and sent to a vision model.\n
  • Boundary markers: The system prompts in scripts/prompt_gen.py lack explicit instructions for the model to ignore or safely handle text embedded within the analyzed visual content.\n
  • Capability inventory: The script has network access, local file read capabilities (for images), and local file write capabilities (for saving results).\n
  • Sanitization: The script performs basic validation on image file extensions and sizes.\n- [DATA_EXFILTRATION]: The skill transmits user-provided image and video data to open.bigmodel.cn. This is the official API endpoint for the Zhipu AI service and is necessary for the skill's intended function.\n- [CREDENTIALS_UNSAFE]: The skill correctly uses the ZHIPU_API_KEY environment variable for authentication, which is a secure practice that avoids hardcoding sensitive credentials.\n- [COMMAND_EXECUTION]: The skill invokes a Python script (scripts/prompt_gen.py) to perform its analysis and API communication, which is the expected behavior for its functionality.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 12, 2026, 06:38 AM
Security Audit — agent-trust-hub — glmv-prompt-gen