The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external CSV and Excel files, creating a vulnerability surface where malicious instructions could be embedded in data. 1. Ingestion points: The _extract_csv and _extract_excel methods in SKILL.md. 2. Boundary markers: Absent; the skill does not use delimiters or warnings to ignore instructions within the data. 3. Capability inventory: The skill has filesystem read/write capabilities via the pandas library and the ability to generate Python scripts. 4. Sanitization: Absent; the implementation does not perform sanitization or validation of the data content prior to processing.
[DYNAMIC_EXECUTION]: The generate_airflow_dag method programmatically constructs Python code by interpolating workflow IDs, task IDs, and descriptions directly into f-string templates. This allows for code injection if an attacker-controlled workflow definition contains Python syntax designed to break out of the string templates. Evidence: Implementation of generate_airflow_dag in SKILL.md which maps metadata fields directly to Python identifiers and operator arguments.

workflow-automation