The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PDF files from external sources and converts them into Markdown text for agent consumption. This establishes an attack surface where a malicious document could contain instructions intended to influence the agent's subsequent actions.
Ingestion points: scripts/search_and_download_papers.py (downloads PDFs); scripts/paddleocr_layout_to_markdown.py (converts PDFs to Markdown).
Boundary markers: Not implemented. The skill does not wrap the extracted text in delimiters or include instructions to ignore embedded commands.
Capability inventory: The skill possesses file writing capabilities, network access via the requests library, and subprocess execution via subprocess.run.
Sanitization: No sanitization or filtering of the extracted OCR text is performed before it is added to the library index.
[DATA_EXFILTRATION]: The skill sends the content of local PDF or image files to an external OCR API at https://ndf9hcz2u2w8ucab.aistudio-app.com/layout-parsing. While this is the intended feature for text extraction, it involves transmitting local document content to a third-party platform (Baidu AI Studio).
[COMMAND_EXECUTION]: The orchestration script scripts/ingest_literature_library.py uses subprocess.run to execute local scripts within the skill directory. The commands are constructed using lists and do not invoke a shell, which follows safety best practices, though they are built dynamically from user-provided search queries and output paths.
[EXTERNAL_DOWNLOADS]: The skill fetches metadata and PDF files from several well-known academic and technical services, including arxiv.org, semanticscholar.org, openalex.org, and huggingface.co. These downloads are documented features for building the paper corpus.

literature-pdf-ocr-library