blip-2-vision-language
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documentation includes examples of processing external image and text data using vision-language models, which is an inherent attack surface for indirect prompt injection.
- Ingestion points: The skill demonstrates model inference using
Blip2Processorandload_model_and_preprocesswhich take local image files (e.g., 'photo.jpg') and user-provided text questions as input inSKILL.mdandreferences/advanced-usage.md. - Boundary markers: The provided code samples do not demonstrate the use of boundary markers or instructions to isolate potentially malicious embedded content within the inputs.
- Capability inventory: The skill showcases inference, batch processing, feature extraction, and fine-tuning capabilities. One integration example uses these capabilities to build a searchable index of image features.
- Sanitization: The documentation does not provide examples of sanitizing or validating ingested visual or textual content prior to model processing.
- [EXTERNAL_DOWNLOADS]: The skill provides instructions for downloading pre-trained models and cloning external code repositories.
- Evidence: The documentation specifies fetching model weights from the official Salesforce organization on HuggingFace and cloning the LAVIS repository from Salesforce's GitHub in
SKILL.mdandreferences/troubleshooting.md. These represent standard operations using well-known services.
Audit Metadata