Approved one action, but the workflow attempted another: an analysis of n8n + MCP + Bitrix24

The author analyzes a security flaw in automated workflows where a gap emerges between human approval and actual execution. Experiments using n8n, Groq, MCP, and Bitrix24 revealed that a system might receive an 'approved=true' flag for one action but then attempt to perform an entirely different operation. This creates a critical vulnerability where the link between user-approved parameters and API call parameters is missing. To address this, the author proposes using the 'Action Envelope' concept, alongside implementing additional checks of the system's initial state before task execution and a control verification after data writing. The article emphasizes the importance of an engineering approach to ensuring data integrity and security within AI agents to prevent unauthorized actions in enterprise systems.