The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content (original and rewritten text) to verify preservation, creating a surface for indirect prompt injection where malicious instructions in the content could attempt to influence the agent's verdict.
Ingestion points: The SKILL.md instructions for manual comparison and the scripts/validate.js script both ingest user-supplied text provided in the agent's context or via file arguments.
Boundary markers: Explicit boundary markers or specific warnings to ignore embedded instructions are absent in the SKILL.md manual comparison guidelines.
Capability inventory: The skill's capabilities are highly restricted, limited to returning a verification status (PASS, FAIL, REVIEW) and identifying repair targets. No file system writes, network operations, or dangerous command executions are present in the provided scripts.
Sanitization: The scripts perform normalization (stripping invisible characters, handling homoglyphs) to improve comparison accuracy, but they do not specifically sanitize or filter the content for adversarial instructions.

preservation-verifier