The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill processes untrusted image data via OCR. Malicious text embedded within a receipt could potentially influence an agent if it blindly trusts the extracted 'vendor' or 'item' names as instructions.
Ingestion points: receipt_scanner.py (via cv2.imread and pytesseract.image_to_string).
Boundary markers: Absent in text extraction.
Capability inventory: Local file read and JSON write.
Sanitization: None; output is raw OCR text formatted by regex.
[Unverifiable Dependencies] (SAFE): The dependencies (pytesseract, opencv-python, pandas, pillow) are industry-standard libraries for image processing and data analysis.
[Data Exposure & Exfiltration] (SAFE): No network requests or access to sensitive local directories (e.g., ~/.ssh) were found. The script only interacts with the user-provided input and output paths.

receipt-scanner