The Agent Skills Directory

[COMMAND_EXECUTION]: The script enrich_papers.py utilizes asyncio.create_subprocess_shell to execute a shell command pipeline that pipes curl output through pdftotext into a local Python script (extract_affiliations.py). Constructing shell commands with dynamic strings (f-strings) is a risky practice that can lead to command injection, although the arxiv_id variable is constrained by a regular expression.
[EXTERNAL_DOWNLOADS]: The skill fetches paper metadata and PDF files from export.arxiv.org and huggingface.co. These are well-known and reputable services for scientific research. Specifically, fetch_and_score.py retrieves paper lists, while download_note_images.py downloads figure images or extracts them from PDFs using pdfimages.
[PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection (Category 8) because it processes untrusted paper titles and abstracts from external sources.
Ingestion points: fetch_and_score.py retrieves content from the arXiv and HuggingFace APIs.
Boundary markers: No explicit boundary markers or isolation instructions were found to separate the untrusted data from the agent's internal logic.
Capability inventory: The skill can perform file writes (e.g., updating .history.json and markdown files) and execute shell commands (e.g., curl, pdfimages) as seen in enrich_papers.py and download_note_images.py.
Sanitization: The code performs basic HTML tag stripping and uses regex for extraction, but these measures are insufficient to mitigate sophisticated prompt injection attempts embedded in paper abstracts.

daily-papers