The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources such as CSV files and SQL databases to perform reconciliation and generate reports. This design creates a surface for indirect prompt injection if the source data contains malicious instructions meant to influence the agent during processing.
Ingestion points: Data ingestion occurs in Step 1 of the Workflow through pd.read_csv and pd.read_sql functions in SKILL.md.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to distinguish between data content and operational instructions.
Capability inventory: The skill includes capabilities to write reports to the local file system (open().write(), df.to_csv()) and display transaction samples to the user.
Sanitization: While the skill standardizes data types (e.g., pd.to_numeric), it lacks sanitization mechanisms to filter or escape potential prompt injection strings within the datasets.

metric-reconciliation