The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied CSV signals and JSON manifests. Risk is mitigated by strict structural and numeric validation in scripts/_common.py, which enforces finite and numeric data types, effectively preventing non-numeric injection via data files.
Ingestion points: scripts/inspect_signal.py, scripts/ecg_hrv_pipeline.py, scripts/eda_pipeline.py, scripts/plan_epochs.py, and scripts/validate_multimodal.py.
Boundary markers: The skill documentation provides explicit warnings regarding deidentification and research scope boundaries.
Capability inventory: Local bounded file writes via atomic_write_bytes and terminal output.
Sanitization: read_numeric_columns ensures all signal data is strictly numeric and finite.
[DATA_EXFILTRATION]: The skill implements strong I/O containment. The file helpers in scripts/_common.py enforce a local root directory, explicitly reject all network-based URLs, and prevent directory traversal through symbolic link rejection.
[PRIVILEGE_ESCALATION]: The skill follows security best practices for data persistence by using os.chmod to restrict file permissions to owner-only (0o600) for all temporary and final output artifacts.

neurokit2