skills/brycewang-stanford/auto-empirical-research-skills/Full-empirical-analysis-skill-R/Gen Agent Trust Hub
Full-empirical-analysis-skill-R
Pass
Audited by Gen Agent Trust Hub on Jul 30, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs the user to install a large number of R packages from the CRAN repository and specific packages from GitHub (e.g., 'lihualei71/cfcausal').
- [REMOTE_CODE_EXECUTION]: The use of 'devtools::install_github' facilitates the download and execution of code from a remote third-party repository.
- [COMMAND_EXECUTION]: The skill includes shell commands for project setup (copying and symlinking files) and R commands for package installation.
- [PROMPT_INJECTION]: Authoritative language such as '必须遵守' (must follow) and 'required' is used to enforce specific output formats and reporting standards.
- [PROMPT_INJECTION]: The skill processes untrusted data files (CSV, Stata, Excel, Parquet) without explicit boundary markers or sanitization, while maintaining file-writing capabilities (e.g., 'ggsave', 'write_json', 'saveRDS'), which constitutes an indirect prompt injection attack surface. Evidence chain for this surface: (1) Ingestion points: 'read_dta', 'read_csv', 'read_excel', 'read_parquet', 'read_sas', 'read_sav', 'stream_in' in 'references/01-data-cleaning.md'; (2) Boundary markers: Absent; (3) Capability inventory: 'ggsave', 'gtsave', 'write_json', 'write_xlsx', 'save_kable', 'save_as_docx', 'saveRDS', 'writeLines'; (4) Sanitization: Absent.
Audit Metadata