The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches PDF documents and metadata from the official arXiv repository (arxiv.org and export.arxiv.org), which is a well-known academic service.
[COMMAND_EXECUTION]: The script constructs file paths using values read directly from arxiv_ids.txt without sanitizing for path traversal characters like ../. A malicious entry in this input file can lead to arbitrary file writes or overwrites of other files within the workspace (e.g., SKILL.md or .gitignore).
[DYNAMIC_EXECUTION]: Uses xml.etree.ElementTree.fromstring to parse API responses from arXiv. This standard library parser is susceptible to XML External Entity (XXE) attacks if the source data were to be compromised, potentially leading to local file disclosure or server-side request forgery.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Reads external PDF content from arXiv and stores extracted headings in a local markdown report (STYLE_REPORT.md).
Boundary markers: The generated report lacks explicit boundary markers or instructions to downstream agents to ignore embedded text in the extracted headings.
Capability inventory: The skill has the capability to perform network requests (urllib) and write to the local file system.
Sanitization: Employs basic regex-based heuristics to identify headings, but these do not sanitize content for potentially malicious instructions that could influence an agent reading the report.

agent-survey-corpus