The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements a model loading pattern in references/chunking-strategies.md that enables the execution of remote repository code.
Evidence: The LateChunker class uses AutoModel.from_pretrained(model_name, trust_remote_code=True) to load embeddings from Jina AI's HuggingFace repository.
This configuration enables the execution of Python code included in the remote model repository during the loading process, which is a significant dynamic execution vector.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture for processing external data is vulnerable to indirect prompt injection attacks.
Ingestion points: Document content is ingested via functions like RecursiveCharacterTextSplitter in SKILL.md and various retrieval functions in references/retrieval-optimization.md.
Boundary markers: The retrieval and generation examples lack explicit delimiters or instruction-isolation markers when incorporating retrieved content into LLM prompts.
Capability inventory: The skill utilizes multiple network-capable clients including openai, cohere, and voyageai, providing a path for data exfiltration if the agent is manipulated by instructions embedded in the documents it processes.
Sanitization: There is no evidence of content sanitization or validation logic in the code snippets provided to identify or block malicious instructions within ingested document text.

rag-architect