fandhe-vector-db
Pass
Audited by Gen Agent Trust Hub on Sep 11, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCREDENTIALS_UNSAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documents components that ingest untrusted data into LLM prompts for query expansion and dictionary extraction (references/search/query-planner.md, references/extensions/dictionary.md). This creates a surface for indirect prompt injection.\n
- Ingestion points: User-provided natural language questions and indexed file content processed into a dictionary context (references/search/query-planner.md).\n
- Boundary markers: Documentation mentions prompt rendering logic but does not specify robust delimiters for the context-to-prompt interpolation in the provided snippets.\n
- Capability inventory: The system makes external API calls to LLM backends (Ollama) and performs data retrieval/indexing operations.\n
- Sanitization: Includes functions like sanitize_question that remove control characters and enforce length constraints.\n- [CREDENTIALS_UNSAFE]: Several sample files in the samples/ directory (e.g., connect-psql.md, connect-node-pg.md, connect-psycopg.md) include hardcoded credentials such as password: "pw-alice". While these are documented as test fixtures from the upstream repository and accompanied by warnings against production use, they remain statically defined secrets in the skill files.
Audit Metadata