smolvlm
Pass
Audited by Gen Agent Trust Hub on Oct 10, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Downloads the SmolVLM model weights from the Hugging Face repository (HuggingFaceTB/SmolVLM-Instruct) during the first run.
- [COMMAND_EXECUTION]: Provides a local Python utility for image analysis.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted image data and prompts, creating a potential surface for indirect prompt injection.
- Ingestion points: The
scripts/view_image.pyscript accepts external image files and text prompts as input. - Boundary markers: Absent; there are no delimiters to separate instructions from untrusted data.
- Capability inventory: Within
scripts/view_image.py, the generated output is limited to printing to stdout, with no hazardous capabilities such as file writing or network activity. - Sanitization: Absent; the input is not sanitized before processing by the vision model.
Audit Metadata