The Agent Skills Directory

[PROMPT_INJECTION]: The file references/hook-compatibility.md provides detailed instructions on how to bypass 'doc blocker' hooks configured in the agent environment. It explicitly recommends using Bash heredoc (cat << 'EOF') to circumvent security interceptions that would otherwise prevent the agent from creating or modifying markdown files.
[INDIRECT_PROMPT_INJECTION]: The skill has a large attack surface for indirect prompt injection (Category 8). It is designed to ingest untrusted data from multiple external sources and process them using the LLM's capabilities.
Ingestion points: The skill uses WebSearch and WebFetch to retrieve content from PubMed, Google Scholar, and journal websites (documented in SKILL.md and references/pubmed-query-builder.md). It also reads user-provided CSV and Excel files for statistical analysis.
Boundary markers: There are no explicit delimiters or instructions to ignore embedded commands when the agent processes the contents of searched literature or data dictionaries.
Capability inventory: The skill possesses extensive capabilities including file system writing (creating project directories and scripts), subprocess execution (pip install, pandoc, pdflatex), and network operations via built-in tools.
Sanitization: No evidence of sanitization or escaping of external content before it is interpolated into prompts or used in script generation.
[COMMAND_EXECUTION]: The statistical analysis scripts (e.g., scripts/analysis-template.py) use dynamic string interpolation to construct formulas for the statsmodels library based on user-provided column names. While intended for legitimate analysis, this pattern could be exploited if malicious data is provided in the input CSV headers.

paper-writer