The Agent Skills Directory

Indirect Prompt Injection (MEDIUM): The prompt templates in references/config_template.yaml use direct interpolation for external data such as {data_samples} and {sample_text} which creates a significant attack surface.
Ingestion points: Data is ingested from external JSON files (train.json, val.json, test.json) and literature PDFs (papers/ directory).
Boundary markers: There are no delimiters (e.g., XML tags or triple quotes) or specific instructions for the agent to ignore instructions embedded within the data samples.
Capability inventory: The system uses these inputs to drive the primary reasoning logic of the agent, including hypothesis generation and data classification (inference).
Sanitization: No evidence of input filtering, escaping, or schema validation is present in the configuration to prevent malicious instructions in the data from overriding the system prompt.

hypogenic