The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external documents to build a persistent knowledge graph, which introduces a surface for indirect prompt injection. Malicious instructions inside documents could be extracted as facts or summaries and later influence the agent's behavior during querying.
Ingestion Points: Untrusted external documents are processed via the EXTRACTION_PROMPT in references/pipeline.md.
Boundary Markers: The skill documentation in references/agent-access.md explicitly recommends serving graph content to agents as "quoted evidence" rather than bare prose to help the agent distinguish between data and instructions.
Capability Inventory: The agent utilizes tools like add_facts to modify the graph and lookup_entity / search_communities to retrieve data, which is then used as context for the QUERY_PROMPT in references/pipeline.md.
Sanitization: The skill advocates for tool-level instructions that define graph content as data rather than instructions, though it lacks programmatic sanitization of the document text itself.

knowledge-graph