The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from multiple external sources, creating an attack surface where malicious instructions could influence agent behavior.
Ingestion points: Reads user-uploaded LaTeX templates from .orx/latex-templates/, processes external run logs via the orx-logs skill, and accepts potential co-author edits synced from Overleaf.
Boundary markers: The instructions do not specify any delimiters or warnings for the agent to ignore embedded instructions within these data sources.
Capability inventory: The skill performs shell command execution (ls, cp), writes to the local file system (paper.tex), and triggers LaTeX compilation (pdflatex, xelatex, lualatex).
Sanitization: There is no mention of sanitization, validation, or escaping of the content retrieved from templates, logs, or external edits before processing.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands (ls and cp) to manage LaTeX templates.
The use of shell globbing (e.g., *.cls, *.sty) and variables in paths (e.g., <name>, <entry>) presents a risk if template directory contents or names contain shell metacharacters or are crafted to exploit path traversal vulnerabilities.
[DYNAMIC_EXECUTION]: The skill triggers the compilation of .tex files using engines like pdflatex, xelatex, or lualatex.
LaTeX is a Turing-complete language that can interact with the file system (e.g., via \input, \read, or \write18 if enabled). Compiling unvetted templates or content from external sources constitutes dynamic code execution.
[EXTERNAL_DOWNLOADS]: The skill references a repository for the alphaXiv OpenResearch project.
The instruction directs the agent to include a citation and link to https://github.com/alphaXiv/OpenResearch in the bibliography of new papers. This is an official vendor resource for this skill.

orx-paper