The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted paper data from external sources and uses it in downstream tasks.
Ingestion points: Fetches paper overviews and markdown from alphaxiv.org and LaTeX source archives from arxiv.org (SKILL.md, Steps 2-4).
Boundary markers: The instructions do not define any delimiters or safety warnings for the agent when processing the fetched text.
Capability inventory: The skill has access to Bash(*), Read, Write, and Glob tools, enabling shell execution and file system modifications.
Sanitization: There is no mention of sanitizing or escaping the content retrieved from AlphaXiv before processing it or passing it to other tools.
[COMMAND_EXECUTION]: The skill executes a bash block that incorporates data fetched from the internet.
Evidence: Step 6 (Research Wiki Ingest) instructs the agent to run a shell command: python3 "$WIKI_SCRIPT" ingest_paper ... --thesis "<one-line thesis from the Tier 1 overview>".
Risk: The <one-line thesis> is derived directly from the external AlphaXiv overview. If the content of a paper's AlphaXiv overview contains shell metacharacters, it could lead to command injection when the agent executes the bash block.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to retrieve paper content from established academic domains.
Evidence: Uses curl to download markdown from alphaxiv.org and .tar.gz source files from arxiv.org.

alphaxiv