attach-image
Pass
Audited by Gen Agent Trust Hub on Sep 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill performs extensive validation on input paths and file contents, including file size checks (max 20MB) and pixel count limits (max 36MP) to prevent resource exhaustion.
- [SAFE]: It employs a defense-in-depth approach to file type verification, checking both file magic bytes and attempting a library-level load with Pillow to confirm the file is a valid image (PNG, JPEG, GIF, or WebP) before processing.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied image files which could contain embedded instructions (e.g., in OCR text). The risk is mitigated by the skill's use of the Pillow library to re-encode and normalize images before they are presented to the model's vision system.
- Ingestion points: User-provided file paths in
src/attach_image/attach_image.py. - Boundary markers: The image is emitted as a structured multimodal attachment via a specific vendor MIME type.
- Capability inventory: Local file reading and IPython display for multimodal context injection.
- Sanitization: Validation of image headers, pixel limits, and re-encoding to JPEG via the Pillow library.
Audit Metadata