The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/render-mermaid.sh script is susceptible to shell command injection. The system() function within the awk block executes a shell command string constructed by concatenating variables, including outdir, which is derived from the user-specified output path. If a directory path contains shell metacharacters like semicolons or backticks, they will be interpreted and executed by the shell.
[COMMAND_EXECUTION]: The scripts/md-to-pdf.sh script has a command injection vulnerability in its construction of the TYPST_HEADER variable. By using an unquoted here-document (<<EOF), the script allows the shell to expand variables such as LINK_COLOR and HEADING_COLOR before they are assigned. An attacker who can control these parameters (e.g., through the --link-color command-line flag) can trigger arbitrary shell command execution.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes user-supplied markdown files using tools like pandoc and mermaid-cli. Maliciously crafted markdown content could exploit potential vulnerabilities in the rendering engines or the provided Lua filters. For instance, the color-spans.lua filter lacks validation when interpolating the color attribute into the generated Typst code.
[EXTERNAL_DOWNLOADS]: The documentation encourages users to manually download a LaTeX template (eisvogel.latex) from a third-party GitHub repository (Wandmalfarbe/pandoc-latex-template). This creates a dependency on external code that is not managed or verified by the skill itself.

pdf