The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface when processing user-uploaded images.
Ingestion points: Text extracted via OCR from images of prescriptions and medical records is stored in the context variable in lib/drug_indication.py.
Boundary markers: The DRUG_INDICATION_PROMPT in lib/drug_indication.py interpolates the extracted text directly into the prompt without utilizing delimiters or specific instructions to the model to ignore any adversarial instructions embedded within the extracted text.
Capability inventory: The skill's capabilities are limited to generating a structured medical review report. It does not have access to tools that perform arbitrary command execution or non-vendor network operations.
Sanitization: No sanitization or escaping is performed on the text extracted from the images before interpolation into the final prompt.
[COMMAND_EXECUTION]: The skill accesses local files (images) using the file:// protocol in lib/llm_client.py to send them to the DashScope vision API. This behavior is restricted by the validate_image_paths utility in lib/image_utils.py, which ensures files exist and have valid image extensions (.jpg, .png, etc.), aligning with the skill's primary function of document analysis.

insurance-claim-medical-review