The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The visualization_agent and statistical_visualization_standards.md guide generate and execute Python (matplotlib/seaborn) or R (ggplot2) code to create research figures. This dynamic code generation and execution process incorporates data extracted from user-provided manuscripts or datasets, representing a potential remote code execution vector if not properly sandboxed.
[COMMAND_EXECUTION]: The formatter_agent and related templates utilize shell commands, including pandoc and tectonic, to convert Markdown documents into LaTeX, DOCX, and PDF formats. These operations involve direct interaction with the system shell and external software binaries.
[PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection by ingesting and processing unstructured data from untrusted external sources.
Ingestion points: revision_coach_agent.md parses reviewer comments from free-form text or PDF pastes; intake_agent.md (Step 10) ingests past writing samples from the user for style calibration; visualization_agent.md processes external datasets.
Boundary markers: The instructions do not mandate the use of delimiters or 'ignore embedded instructions' warnings when handling these external texts.
Capability inventory: The pipeline's existing capabilities for code and command execution increase the risk associated with successful prompt injection.
Sanitization: There is no evidence of sanitization or input validation performed on external content before it is interpolated into the agent context.

academic-paper