The Agent Skills Directory

[Data Exposure & Exfiltration] (SAFE): The code snippets use local or placeholder file paths (e.g., './docs', 'documents.txt') and do not contain hardcoded credentials or unauthorized network exfiltration logic. Standard use of SerpApi is noted but requires user-provided keys.
[Indirect Prompt Injection] (SAFE): The skill demonstrates patterns for Retrieval-Augmented Generation (RAG) and agent-based tool use (e.g., 'send_email', 'search_database'), which are potential surfaces for indirect prompt injection. However, these are presented as educational templates without malicious intent.
Ingestion points: sub-skills/pattern-1-rag-with-langchain.md (reads documents.txt), sub-skills/2-batch-processing.md (reads ./docs).
Boundary markers: Absent (standard for architectural templates).
Capability inventory: Agent tools in sub-skills/5-callbacks.md (search/math) and sub-skills/pattern-2-custom-agent-with-tools.md (email/db placeholders).
Sanitization: Not implemented in snippets.
[Unverifiable Dependencies & Remote Code Execution] (SAFE): The skill imports standard, well-known libraries such as langchain, pytest, and concurrent.futures. There are no commands for remote script downloads (e.g., curl|bash) or dynamic execution of untrusted code.

langchain-architecture