When LLMs Refuse, But Agents Execute: The Hidden Tool Layer Vulnerability

Testing a real LangGraph AI agent reveals a critical security gap: while the LLM successfully recognizes dangerous inputs like SQL injections and path traversals, the tool execution layer blindly trusts these decisions and executes them anyway. This vulnerability exposes that LLM-level safety testing is insufficient—developers must implement validation and sanitization at the tool layer, not rely solely on the model's judgment.