The Agent Skills Directory

[PROMPT_INJECTION]: The skill demonstrates application architectures that are susceptible to indirect prompt injection due to the way external data is interpolated into instructions.
Ingestion points: The documentation describes patterns using TextLoader to ingest local documents and SequentialChain to process arbitrary text input variables.
Boundary markers: The provided PromptTemplate examples (e.g., template="Extract key entities from: {text}\n\nEntities:") directly interpolate content into instructions without using delimiters such as XML tags, triple quotes, or explicit 'ignore embedded instructions' warnings.
Capability inventory: The skill documents the creation of tools with significant capabilities, including database searching (search_database) and email dispatching (send_email), which could be abused if an injection occurs.
Sanitization: There is no mention or demonstration of input validation, escaping, or content filtering for the data being passed into the LLM components.

langchain-architecture