The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The LangChain integration example in 'SKILL.md' includes the configuration allow_dangerous_deserialization=True when calling FAISS.load_local(). This parameter bypasses security checks and allows the underlying pickle library to execute arbitrary Python code that may be embedded in a malicious index file.
[COMMAND_EXECUTION]: By encouraging the use of unsafe deserialization flags, the skill creates a path for system command execution if the agent is instructed to load a vector index from an untrusted or compromised source.
[PROMPT_INJECTION]: The skill describes an attack surface for indirect prompt injection by ingesting untrusted vector data into the agent's context without adequate sanitization.
Ingestion points: 'SKILL.md' and 'references/index_types.md' use functions like faiss.read_index and FAISS.load_local to import data from external files.
Boundary markers: There are no instructions or examples defining delimiters to separate ingested data from agent instructions.
Capability inventory: The skill has the capability to read from the file system and perform similarity searches that influence agent responses.
Sanitization: No sanitization logic is provided, and security warnings for deserialization are explicitly disabled in the code examples.

faiss