The Agent Skills Directory

[SAFE]: The skill is a legitimate research tool with no detected security issues. It implements well-known academic workflows and encourages security best practices for authors (such as anonymization and citation verification).
[DATA_EXPOSURE_EXFILTRATION]: The skill uses academic APIs (Semantic Scholar, arXiv, CrossRef) for literature discovery and citation verification. These are well-known, trusted services in the research community. Local file access is limited to standard project discovery (e.g., searching for results or configurations within the workspace).
[REMOTE_CODE_EXECUTION]: The skill includes a Makefile that downloads official LaTeX style files from the NeurIPS conference servers, which is a standard procedure for maintaining template compliance. It also references the Exa MCP tool for academic search via npx, a well-known service.
[PROMPT_INJECTION]: The skill contains role-play instructions (e.g., 'You are an expert reviewer') used to guide sub-agents during iterative refinement phases. These are internal logic patterns for the 'Autoreason' methodology and do not attempt to override the parent agent's safety protocols or system instructions.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an inherent attack surface as it ingests external research papers from the web. While a malicious paper could theoretically contain adversarial instructions, the skill is focused on extracting citations and summarizing findings from reputable sources like arXiv, which is a standard and generally acceptable research task. No vulnerabilities in the handling of this data were identified.

research-paper-writing