The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a Graph RAG (Retrieval-Augmented Generation) architecture that creates an indirect prompt injection vulnerability surface by design.\n
Ingestion points: Untrusted data from conversations and business logs enters the agent context through the add_episode and add_memory methods documented in SKILL.md.\n
Boundary markers: The 'Context Assembly' code example in SKILL.md demonstrates direct interpolation of retrieved facts into the system_prompt string without the use of delimiters or 'ignore' instructions to prevent the LLM from obeying embedded commands.\n
Capability inventory: The skill facilitates the integration of this context with LLMs (Claude 3.5) and action-oriented frameworks like CrewAI and LangGraph, as seen in the integration examples in SKILL.md.\n
Sanitization: While SKILL.md provides a redact_pii function for privacy, it lacks instructions or patterns for sanitizing, escaping, or validating the content of retrieved facts to prevent prompt injection.

graphiti