The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external dataset text and model-generated hypotheses. It mitigates injection risks by treating all such text as untrusted data, applying strict schema validation, and redacting raw content in audit reports and inspections.
Ingestion points: scripts/audit_dataset.py (dataset JSON), scripts/inspect_outputs.py (hypothesis bank JSON), and scripts/evaluate_local.py (result JSON).
Boundary markers: The SKILL.md and example task configuration files include explicit instructions to treat all supplied text as quoted data and not as instructions.
Capability inventory: The bundled tools perform file I/O, hashing, and numeric analysis. They do not have the capability to execute instructions found within the data.
Sanitization: The skill uses strict type-checking, length limits, and numeric bounds (e.g., node counts, document depth) to prevent malicious payloads from influencing agent behavior.
[COMMAND_EXECUTION]: The skill uses uv and python3 to run local, deterministic scripts for auditing and planning. These scripts do not interpolate untrusted text into shell commands or dynamically execute code.
[EXTERNAL_DOWNLOADS]: The skill references the hypogenic package and official ChicagoHAI repositories. It provides specific SHA-256 hashes for the package artifacts and immutable commit SHAs for datasets, ensuring supply chain integrity through verification.
[CREDENTIALS_UNSAFE]: The skill implements active detection for hardcoded secrets. It rejects configuration files containing common API key patterns or field names like api_key or secret. It checks for the presence of environment variables using specialized helper functions that only return presence as a boolean without exposing the variable's value.
[SAFE]: The YAML parser in scripts/_common.py uses a SafeLoader subclass and performs a manual token scan to reject aliases, anchors, and explicit tags, effectively mitigating YAML-based resource exhaustion or object injection attacks. This usage addresses the security concerns typically associated with standard YAML loading.

hypogenic