The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external URLs (conference submission links) using the WebFetch tool to generate a README.md and determine submission requirements. This creates an attack surface where a malicious website could contain instructions designed to influence the agent's behavior, such as altering the cleanup logic or injecting malicious text into the output files.
Ingestion points: SKILL.md (Step 6: Generate README)
Boundary markers: None mentioned; the skill does not instruct the agent to ignore instructions embedded in the fetched web content.
Capability inventory: The skill has file-write, directory creation, and shell execution capabilities.
Sanitization: None specified for the web content before interpolation into the README template.
[COMMAND_EXECUTION]: The skill executes shell commands to process files extracted from user-provided .zip archives. Several patterns used in the bash blocks are susceptible to exploitation if the zip file contains filenames with shell metacharacters or directory traversal sequences:
Evidence: The use of while read f; do cp "$f" output/tables/; done in Step 5 (Handle Images and Tables) can be manipulated if filenames contain newlines or other shell-sensitive characters.
Evidence: The command find /tmp/latex-template-temp -type f ... -exec cp {} output/styles/ \; is safer than a pipe to while read, but processing files in a temporary directory derived from an untrusted zip file still carries a risk of path traversal if not handled by the agent's underlying shell executor guardrails.

latex-conference-template-organizer