The Agent Skills Directory

[COMMAND_EXECUTION]: Potential SQL injection vulnerability in the PostgresHybridSearch template.\n
Evidence: In SKILL.md, the hybrid_search method constructs a SQL query by interpolating dictionary keys directly into the where_clause string: where_clause += f" AND metadata->>'{key}' = ${len(params)}". While values are parameterized, the keys are not, allowing an attacker who controls the metadata keys to manipulate the SQL logic.\n- [EXTERNAL_DOWNLOADS]: Fetches pre-trained models from a trusted well-known service.\n
Evidence: The search_with_rerank method in SKILL.md uses sentence_transformers to load cross-encoder/ms-marco-MiniLM-L-6-v2. This operation downloads model weights from Hugging Face's official repository.\n- [PROMPT_INJECTION]: Exposure to indirect prompt injection through search query processing.\n
Evidence Chain:\n
Ingestion points: The query string enters the system via the hybrid_search and search methods in the templates.\n
Boundary markers: Absent; the query is passed directly to database search functions and neural reranking models without delimiters.\n
Capability inventory: Executes database queries (PostgreSQL/Elasticsearch) and performs model inference which can influence RAG output.\n
Sanitization: Database values are parameterized, but the raw query text is processed by semantic search engines without specific filtering for embedded instructions.

hybrid-search-implementation