The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data (CSV, Excel, JSON), creating a surface for indirect prompt injection attacks where malicious instructions in the data could influence agent behavior.
Ingestion points: Data is loaded into the agent's context using pandas reading functions in both SKILL.md and the scripts/analyze.py tool.
Boundary markers: The instructions and script lack explicit boundary markers or instructions to treat data as untrusted content when printing summaries, column names, or values.
Capability inventory: The skill has capabilities to read/write files and generate statistical reports.
Sanitization: The generate_report function in scripts/analyze.py interpolates raw data values directly into HTML templates without sanitization, allowing potential data poisoning to persist in generated outputs.
[DYNAMIC_EXECUTION]: The skill demonstrates the use of the pickle module for data persistence, which is a known risk factor.
Evidence: SKILL.md provides an example of using pickle.dump to export analysis results. While only serialization is shown, the inclusion of pickle patterns facilitates unsafe deserialization practices if the agent or user later loads these files.

data-analysis