blip-2-vision-language

Pass

Audited by Gen Agent Trust Hub on Mar 28, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: Fetches pre-trained model weights from Salesforce's official HuggingFace repository.\n- [EXTERNAL_DOWNLOADS]: Recommends installation of standard machine learning libraries from official package registries.\n- [COMMAND_EXECUTION]: Implements local model inference, training, and deployment workflows using standard Python machine learning frameworks.\n- [DATA_EXFILTRATION]: Provides a troubleshooting helper function that performs network requests to external URLs using the requests library.\n- [PROMPT_INJECTION]: Ingests external image data and user questions, which serves as a surface for indirect prompt injection attacks.\n
  • Ingestion points: image_path and question inputs in multiple code snippets (e.g., VisualQA, ImageCaptionTool, ImageSearchEngine).\n
  • Boundary markers: Absent; inputs are passed directly to processing functions without delimiters or isolation instructions.\n
  • Capability inventory: File system access via PIL.Image.open and LLM inference via model.generate.\n
  • Sanitization: No input validation or sanitization is demonstrated in the provided examples.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 28, 2026, 06:07 PM