The Agent Skills Directory

[COMMAND_EXECUTION]: Several code patterns demonstrate unsafe Cypher query construction for Neo4j. In Pattern 1 (SKILL.md), relationship types and entity labels are concatenated into strings. In Pattern 4 (SKILL.md), entity names extracted from user queries are directly interpolated into the query string (MATCH (e {name: '{entity}'})) instead of using parameterized queries. If these variables are populated by untrusted LLM-extracted data, it allows for Cypher injection attacks against the underlying graph database.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture is designed to process untrusted external data (documents and user queries) to populate and query a knowledge graph, creating a significant attack surface.
Ingestion points: Document text ingestion in the extract_knowledge function (SKILL.md) and user-provided queries in the HybridGraphRAG.retrieve method (SKILL.md).
Boundary markers: The prompt templates, such as EXTRACTION_PROMPT, lack boundary markers or explicit instructions to ignore malicious commands embedded within the text content.
Capability inventory: The skill interacts with Neo4j databases via network protocols and performs indexing operations that involve file system writes through the graphrag package.
Sanitization: There is a lack of robust input sanitization or validation for the entities and relationships extracted from the processed text before they are used to update the database state or execute queries.

graphrag-patterns