The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill documentation and example scripts include various hardcoded default credentials meant for development and testing.
SKILL.md contains the placeholder password in a connection initialization example: db = GraphDB("bolt://localhost:7687", "neo4j", "password").
references/neo4j.md includes hardcoded credentials in Docker setup instructions: NEO4J_AUTH=neo4j/password.
references/arangodb.md features hardcoded root passwords in Docker and driver examples: ARANGO_ROOT_PASSWORD=password and auth: { username: 'root', password: 'password' }.
examples/knowledge-graph/hybrid_vector_graph.py provides a default password in the class initialization: neo4j_password="password".
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of externally controlled database records into LLM prompts without adequate safety markers or sanitization.
Ingestion points: Data is sourced from the Neo4j database (specifically Concept node descriptions) via the hybrid_search and get_context_for_llm methods in examples/knowledge-graph/hybrid_vector_graph.py.
Boundary markers: The context generator uses Markdown headers to structure output but fails to include explicit delimiters or instructions telling the model to ignore potential commands embedded within the individual record descriptions.
Capability inventory: The skill provides methods to execute Cypher queries and aggregate data into context strings for AI chat completion.
Sanitization: There is no logic implemented to sanitize or escape retrieved database content before it is interpolated into the prompt context string, allowing any malicious instructions stored in the database to be interpreted by the downstream agent.

using-graph-databases