The Agent Skills Directory

[DYNAMIC_EXECUTION]: The PgVectorStore.search method in resources/implementation-playbook.md is vulnerable to SQL injection through unsafe string concatenation.
Evidence: The code constructs a SQL query by interpolating keys from the filter_metadata dictionary directly into the string: f"metadata->>'{key}' = ${len(params)}".
Risk: If an application using this template allows untrusted input to influence the keys of the metadata filter, an attacker could execute arbitrary SQL commands to leak data or bypass security controls.
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to facilitate Retrieval-Augmented Generation (RAG), which introduces a vulnerability surface for indirect prompt injection.
Ingestion points: Untrusted data is retrieved from external vector databases via search methods in resources/implementation-playbook.md (e.g., PineconeVectorStore.search, QdrantVectorStore.search).
Boundary markers: Absent. The templates return raw strings and dictionary objects from the database without any delimiters or warnings to ignore embedded instructions.
Capability inventory: The skill templates include network-based API calls to Pinecone, Qdrant, and Weaviate, as well as database operations for PostgreSQL.
Sanitization: Absent. The provided implementation patterns do not include any logic to sanitize, escape, or validate retrieved content before it is likely used in a downstream LLM prompt.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The implementation playbook includes code that downloads external model assets at runtime.
Evidence: CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2') in the _rerank method of the PineconeVectorStore class.
Source: Fetches model weights and configuration from Hugging Face, a well-known and established service for machine learning models.

similarity-search-patterns