The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from untrusted documents (images, PDFs, and Office files) that could contain adversarial instructions targeting the agent.
Ingestion points: External files are read via file_path and file_url arguments in scripts/client.py and scripts/unlimited_ocr_caller.py.
Boundary markers: The extracted text is returned in a structured JSON envelope, but the skill body provides a clear warning for the agent to treat all extracted text as untrusted data.
Capability inventory: The skill includes network capabilities via the httpx library and file-writing capabilities through the CLI wrapper.
Sanitization: There is no automatic sanitization or filtering of the extracted text content; the skill relies on the calling agent to apply its own safety policies.
[DATA_EXPOSURE]: The skill accesses the local filesystem to read documents for parsing.
Evidence: The source_file function in scripts/client.py resolves local paths and reads file contents into memory to prepare them for transmission to the OCR provider.
[DATA_EXFILTRATION]: The skill transmits document content to external servers to perform OCR tasks.
Evidence: scripts/client.py sends Base64-encoded document data to Baidu Cloud API endpoints (aip.baidubce.com) or a user-configured local backend URL. While this is the intended functionality, it constitutes the movement of local data to remote services.

unlimited-ocr-document-parsing