The Agent Skills Directory

[PROMPT_INJECTION]: The reference file references/query-classification-taxonomy.md contains common prompt injection strings (e.g., "Ignore previous instructions and...") listed within a table of query classes. These are provided as illustrative examples for building query routers and classifiers, and they do not represent an attempt to override the current agent's instructions.
[DATA_EXFILTRATION]: The skill includes extensive guidance on implementing multi-tenancy and data isolation. It emphasizes security best practices, such as using mandatory pre-filters at the database level rather than relying on the LLM for access control, which is a critical measure for preventing cross-tenant data leakage.
[INDIRECT_PROMPT_INJECTION]: The skill documentation correctly identifies the risk of indirect prompt injection in RAG pipelines. It explicitly warns developers to "Treat retrieved content as untrusted — prompt injection in documents is real" in SKILL.md and provides architectural patterns such as strict grounding prompts, structured answer contracts, and citation verification to mitigate these risks.
[EXTERNAL_DOWNLOADS]: The documentation references well-known technology tools and services such as Unstructured.io, LlamaParse, and Vertex AI. These are industry-standard tools for document processing and RAG architecture. No suspicious or untrusted download patterns or scripts were identified.

rag-architect