The Agent Skills Directory

Data Processing (Indirect Prompt Injection): This skill is designed to process untrusted clinical text, which presents an inherent surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through file paths or inline text passed to the scripts/deidentify.py script.
Boundary markers: The skill uses deterministic pseudonym tokens (e.g., [NAME-1], [DATE:1985]) to delimit redacted content and produces a structured audit manifest.
Capability inventory: The redaction script uses the Python standard library only (with an optional check for spacy), lacks network access, and is restricted to reading/writing local files in the agent sandbox.
Sanitization: The skill implements a verification guard that scans the redacted output for any leaked original values before release, flags unknown capitalized word pairs for human review, and hashes removed values in the manifest using SHA-256 with a salt.
Sensitive Artifact Management (Re-identification Crosswalk): The script includes an optional --map feature that generates a reversible token-to-value crosswalk. This artifact is highly sensitive as it enables re-identification of the data. The skill instructions correctly identify this risk and provide clear guidance to the agent to warn users and store the file securely and separately from redacted output.

phi-deidentifier