The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including business specifications and implementation code, during its core workflow (INGEST and EXTRACT phases). This creates a surface for indirect prompt injection.
Ingestion points: Specification documents (markdown, PRD, User Stories) and source code artifacts (e.g., src/controllers/auth.ts).
Boundary markers: The skill employs structured YAML finding blocks and Given/When/Then BDD scenario formats. It explicitly instructs the agent to avoid 'Circular Verification' by not reading expected values from the implementation itself.
Capability inventory: The skill is restricted to static file reading and code searching. It does not have file-write or network-execution capabilities.
Sanitization: The skill mandates 'Adversarial Probing' across six categories (Boundary, Omission, Contradiction, Implicit, Negative, Concurrency) to identify malicious or contradictory instructions in source data.
[DATA_EXFILTRATION]: Automated scanners detected a malicious URL (nodomain.com). However, this domain appears strictly as a non-functional string literal within the reference/fix-prompt-generation.md file. It is used as an example of a malformed email address for a negative test case (e.g., 'add scenarios for "notanemail", "missing@dot", "@nodomain.com"'). There is no code that attempts to resolve or communicate with this domain.
[METADATA_POISONING]: The reference/fix-prompt-generation.md file was flagged by a reputation scanner ('FileRepMalware'). Analysis indicates this is a false positive likely triggered by the presence of the aforementioned blacklisted placeholder domain or the nature of the LLM remediation prompt templates contained in the file. The file is standard markdown documentation for the skill's remediation workflow and contains no executable or malicious logic.

attest