The Agent Skills Directory

Indirect Prompt Injection (LOW): The skill processes untrusted user input and external data across multiple examples in examples.md, creating a surface for indirect prompt injection.
Ingestion points: External data enters through the articles, question, approved_content, documents, data, and sources fields across all defined dspy Signatures.
Boundary markers: The skill uses dspy Signatures to enforce structured input schemas and explicitly instructs the model to use 'only the provided' context in its docstrings.
Capability inventory: No dangerous capabilities such as subprocess calls, file system writes, or unauthorized network operations were found in the scripts.
Sanitization: The skill demonstrates best practices by implementing output sanitization via regex citation validation, secondary LLM faithfulness judges, cross-checking between independent model generations, and confidence-based gating.

ai-stopping-hallucinations