The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill's setup documentation (references/r-guides/00-setup.md) contains instructions to execute system commands with administrative privileges using sudo to install library dependencies on Ubuntu/Debian systems.
[REMOTE_CODE_EXECUTION]: Several guides (references/r-guides/00-setup.md, references/r-guides/03-subgroup-plots.md) recommend installing R packages from third-party GitHub repositories using devtools::install_github() or remotes::install_github(), which involves downloading and executing unverified external code from unknown sources.
[COMMAND_EXECUTION]: The workflow defined in SKILL.md executes an external Python script using uv run tooling/python/session_log.py, representing the execution of logic defined outside the skill's scope.
[EXTERNAL_DOWNLOADS]: The skill performs numerous package downloads from CRAN and GitHub. While CRAN is a well-known repository, the large number of external dependencies increases the attack surface.
[INDIRECT_PROMPT_INJECTION]: The analysis workflow ingests study data from external CSV files and uses it to generate statistical models and figures without boundary markers or sanitization.
Ingestion points: 05_extraction/extraction.csv is read by multiple analysis scripts, including assets/r/02_effect_sizes.R and assets/r/06_tables.R.
Boundary markers: None mentioned for data ingestion.
Capability inventory: Executes R scripts, performs file system writes to results directories, and runs shell commands.
Sanitization: No sanitization or validation of the input CSV data is performed before processing.

ma-meta-analysis