The Agent Skills Directory

[SAFE]: The skill is a defensive utility designed to enhance privacy and security. It provides operational directives for identifying and redacting sensitive information across various surfaces including prompts, logs, and RAG outputs.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data during its privacy review workflows. Ingestion points: Ingests outbound LLM prompts, RAG chunks, and tool results (specified in SKILL.md and gemini.md). Boundary markers: Instructs the agent to implement typed stable placeholders (e.g., <<EMAIL_1>>) to maintain data relationships while masking content. Capability inventory: Performs PII classification and metadata-only audit logging; no direct file-system writes or unauthorized network operations are present. Sanitization: Implements redaction, masking, and hashing (SHA-256) of identified sensitive fields and blocks credentials entirely.
[SAFE]: No evidence of obfuscation, malicious command execution, or remote code execution was detected. The Python code block is an illustrative example of secure, in-memory redaction logic using regular expressions.

ai-pii-redaction-review