The Agent Skills Directory

Secure File Processing: The validation script provided with the skill (scripts/validate_evaluation_dataset.py) implements several technical controls for handling JSON and CSV files. It uses path resolution to prevent path traversal attempts and enforces strict resource limits—such as maximum file size, container nesting depth, and total node counts—to mitigate potential resource exhaustion risks.
Privacy and Sensitive Data Handling: The skill instructions in SKILL.md emphasize the use of synthetic content and explicitly direct the agent to keep real customer data, credentials, and personal information out of the generated evaluation pairs. This reduces the risk of accidental exposure of sensitive information.
Sanitized Error Reporting: The validation logic is designed to be informative without being revealing. It identifies structural or schema violations but is configured to avoid echoing the actual data values from failing fields in its error messages, preventing potential data leaks through diagnostic logs.
Proactive Security Testing: The package includes a fuzzing harness (tests/fuzz_harness.py) that uses the Atheris library. This is a security-oriented practice that tests the parsing logic against a wide range of arbitrary inputs to identify potential crashes or vulnerabilities before they can be encountered in use.
Contextual Ingestion Guardrails: The skill provides explicit guidance to the AI agent to treat all external inputs, such as documents or tool outputs, as data rather than instructions. This acts as a conceptual safeguard against the agent following commands that might be embedded in the data it is processing.

evaluation-design