clean-data
Pass
Audited by Gen Agent Trust Hub on Oct 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes clinical datasets (CSV/Excel) and researcher-provided codebooks to generate analysis reports. This constitutes an ingestion point for untrusted data that could contain malicious instructions.
- Ingestion points: Dataset file paths and data dictionaries processed in SKILL.md (Stage 1 and 2).
- Capability inventory: The skill has access to Bash, Write, Edit, Read, Grep, and Glob tools. It executes internal Python scripts via shell commands as demonstrated in tests/test_reverse_coding.sh and tests/test_structural_zero.sh.
- Boundary markers: The instructions do not define specific delimiters or "ignore instructions" wrappers when the agent processes or summarizes external data.
- Sanitization: No specific filtering or sanitization of data content is mentioned prior to code generation or report building.
- [DYNAMIC_EXECUTION]: The skill generates and facilitates the execution of Python and R code for data cleaning purposes.
- Evidence: SKILL.md Stage 3 defines the generation of cleaning scripts for missing values, outliers, and duplicates using templates like references/profiling_template.py.
- Execution method: The generated code is intended for researcher execution, and regression tests use the Bash tool to run the skill's Python-based detectors.
Audit Metadata