The Agent Skills Directory

[COMMAND_EXECUTION]: The skill directs the agent to execute shell commands such as pdftotext, pandoc, and curl to fetch and process external documents. These commands are executed on files downloaded from arbitrary URLs found via search engines. (File: references/data-collection.md)
[REMOTE_CODE_EXECUTION]: The instructions explicitly permit the agent to modify its environment by installing external software packages using brew install (poppler, pandoc) and pip install (pdfplumber, python-docx, python-pptx) if they are not already present. (File: references/data-collection.md)
[EXTERNAL_DOWNLOADS]: The skill uses curl -L -o to download binary files (PDF, DOCX, PPTX) from unverified external URLs identified during the search phase. (File: references/data-collection.md)
[INDIRECT_PROMPT_INJECTION]: The skill processes large volumes of untrusted data from the web and external files (Phase 1-3). It lacks mandatory boundary markers or specific sanitization instructions to prevent embedded malicious instructions in those files from overriding the agent's primary mission. (File: SKILL.md, references/data-collection.md)
Ingestion points: Web scraping via scrape_as_html and scrape_as_markdown; file reading via converted Markdown from pdftotext and pandoc.
Boundary markers: Absent. The skill does not define specific delimiters or instructions to ignore commands within the gathered data.
Capability inventory: Shell execution for curl, brew, pip, pdftotext, and pandoc.
Sanitization: The skill relies on the agent to "extract key information" but provides no technical sanitization or validation of the content.

competitive-analysis