The Agent Skills Directory

[COMMAND_EXECUTION]: Potential SQL Injection vulnerability in the PgVectorStore.search method.
Evidence: conditions.append(f"metadata->>'{key}' = ${len(params)}") in SKILL.md.
Description: The implementation uses f-string interpolation for database keys within metadata filters. If the keys in the filter_metadata dictionary are derived from untrusted user input, an attacker could manipulate the SQL query structure to bypass access controls or extract sensitive data.
[EXTERNAL_DOWNLOADS]: Fetches pre-trained machine learning models from HuggingFace Hub.
Evidence: model = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2') in SKILL.md.
Description: The Pinecone template includes a reranking method that automatically downloads and loads a cross-encoder model from a well-known external repository at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill manages the lifecycle of external data used for Retrieval Augmented Generation (RAG), which presents a surface for indirect prompt injection.
Ingestion points: upsert methods in Pinecone, Qdrant, pgvector, and Weaviate templates.
Boundary markers: None present in the database interaction templates.
Capability inventory: Database read/write operations and vector similarity searches.
Sanitization: While values are correctly parameterized in the SQL template, the metadata keys are unsanitized, creating a vulnerability surface.

similarity-search-patterns