The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, specifically web search results and academic metadata from APIs like Semantic Scholar. This data is used to populate a knowledge graph and generate an interactive HTML report. The absence of specific boundary markers or robust HTML escaping when interpolating this search data into the report template creates a potential surface for indirect injection or stored cross-site scripting (XSS) if a user opens the generated report.
Ingestion points: Search results from WebSearch/Brave Search and API responses from Semantic Scholar (as processed in SKILL.md and paper-downloader.py).
Boundary markers: None explicitly defined to separate untrusted data from agent instructions during the extraction phase.
Capability inventory: The agent has permissions for file system writes (generating HTML, JSON, and Markdown reports), network downloads (fetching PDF files), and local command execution (spawning the provided Python script).
Sanitization: While the Python script sanitizes filenames to prevent path traversal issues, there is no evidence of content sanitization or escaping for data placed within the learning-guide-template.html.
[COMMAND_EXECUTION]: The skill utilizes an included local Python script, paper-downloader.py, to manage the search and download of academic papers. The agent is instructed to invoke this script with arguments derived from search results (such as titles and IDs). The script itself uses only Python standard libraries and includes logic to sanitize filenames, reducing the risk of unintended file system operations.
[EXTERNAL_DOWNLOADS]: The skill fetches academic papers (PDFs) from external repositories. The primary targets are well-known and trusted academic services, including arXiv (arxiv.org) and Semantic Scholar (semanticscholar.org). These downloads are consistent with the skill's stated purpose of assisting with academic research and learning.

deep-learning