The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documentation for Retrieval Augmented Generation (RAG) and Weaviate Agents demonstrates patterns where untrusted data from database collections is interpolated into LLM prompts without sufficient protection. Specifically, in references/08-rag.md and references/12-agents.md, user-controllable properties like body or abstract are injected into prompts (e.g., single_prompt="Summarize: {title} — {body}"). This creates a surface where malicious content stored in the database could attempt to override agent instructions.
Ingestion points: Text properties retrieved from Weaviate collections such as body, abstract, and review_body (references/08-rag.md, references/12-agents.md).
Boundary markers: The provided code examples lack explicit delimiters or instructions to the LLM to ignore potentially malicious embedded content in the variables.
Capability inventory: The skill enables the agent to interact with various LLM providers and perform data transformations (Transformation Agent), meaning a successful injection could influence these actions.
Sanitization: The reference documentation does not demonstrate sanitization or validation of database content before it is processed by the generative model.
[EXTERNAL_DOWNLOADS]: The included utility script scripts/check-updates.py uses urllib.request to connect to the official PyPI registry (pypi.org) to verify the latest version of the weaviate-client package. This is a standard administrative function using a well-known and trusted service.
[COMMAND_EXECUTION]: The skill provides instructions for standard environment setup, including installing packages via pip and npm, and running database instances via docker and docker compose (SKILL.md, references/00-overview.md).

weaviate