The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates processing external data files using pd.read_csv and glob.glob. This constitutes a theoretical attack surface for indirect prompt injection if an agent were to interpret the data content as instructions.
Ingestion points: pd.read_csv('data.csv') and glob.glob('data/*.csv') in SKILL.md.
Boundary markers: The templates do not specify delimiters or warnings for the agent when processing the content.
Capability inventory: The skill includes file system write operations via to_csv in SKILL.md.
Sanitization: The examples focus on data cleaning for statistical purposes rather than security-focused sanitization of input strings.

data-transform