The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses pandoc and tectonic for document conversion and LaTeX compilation in the formatter_agent. These are standard tools for academic workflows and are used within their intended scope.
[REMOTE_CODE_EXECUTION]: The visualization_agent generates Python (matplotlib, seaborn) and R (ggplot2) code for creating publication-quality charts. This code is intended for data visualization and uses well-known, legitimate libraries.
[EXTERNAL_DOWNLOADS]: The skill references numerous well-known academic organizations, conferences, and databases including Nature, Science, ICLR, NeurIPS, ACL, EMNLP, and HEEACT. All links point to official and reputable domains.
[DATA_EXFILTRATION]: No sensitive file paths (e.g., SSH keys, credentials) are accessed, and no unauthorized network operations were found.
[PROMPT_INJECTION]: The instructions contain 'IRON RULE' and 'COMMITMENT GATE' markers to guide agent behavior. These are self-instructional and do not attempt to bypass system-level safety protocols.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external data such as reviewer comments and past papers for analysis. While this presents a theoretical attack surface, the risk is inherent to the academic revision use-case, and no exploitable pattern was detected.

academic-paper