The Agent Skills Directory

[DYNAMIC_EXECUTION]: The LangChain integration example in SKILL.md uses the allow_dangerous_deserialization=True flag when calling FAISS.load_local. This parameter enables the use of the pickle module to load index metadata. Deserializing data with pickle is insecure and can lead to arbitrary code execution if the index file originates from an untrusted source.
[INDIRECT_PROMPT_INJECTION]: The skill establishes a capability to load and process external vector index files, which serves as an attack surface for indirect injection. By documenting how to bypass deserialization safety checks, the skill creates a path for an attacker to achieve code execution via a crafted index file.
Ingestion points: The FAISS.load_local("faiss_index", ...) function call in SKILL.md which ingests files from the local file system.
Boundary markers: No delimiters or warnings regarding untrusted index files are present in the instructions.
Capability inventory: The skill utilizes faiss and numpy for vector search, and uses faiss.read_index, faiss.write_index, and LangChain's local storage methods for file I/O.
Sanitization: No sanitization or integrity checks are performed on the index data before loading it with the dangerous flag enabled.
[EXTERNAL_DOWNLOADS]: The skill documentation includes standard pip install commands for the faiss-cpu and faiss-gpu packages from official package registries.

faiss