The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides numerous architectural patterns and code examples for building RAG and Agent systems that ingest untrusted external data (such as web search results or database records) and interpolate them directly into LLM prompts.
Ingestion points: The rag_chain.invoke() and agent_executor.invoke() patterns in SKILL.md and EXAMPLES.md demonstrate the ingestion of arbitrary text into the agent's context.
Boundary markers: While some examples use <context> XML-style tags, many rely on simple f-string or template interpolation without robust delimiters to distinguish between system instructions and retrieved data.
Capability inventory: The skill documentation includes definitions for tools that perform mathematical calculations, data analysis, and simulated web searching.
Sanitization: The provided code snippets do not implement input validation, content filtering, or escaping techniques to mitigate the risk of instructions embedded in the retrieved data influencing the agent's behavior.
[DYNAMIC_EXECUTION]: The documentation includes implementation examples for a 'Calculator' tool that utilizes the Python eval() function to process user-supplied mathematical expressions.
Evidence: SKILL.md contains a calculator_tool implementation using eval(expression). EXAMPLES.md contains a calculate function using eval(expression, {"__builtins__": {}}, {}). While the latter attempts to restrict the environment, both patterns represent dynamic execution of strings that may originate from untrusted LLM tool calls.

langchain-orchestration