The Approval-Runtime Gap: How AI Agent Tools Quietly Escalate Risk

As AI agents increasingly control real systems through tools that read files, send emails, and access databases, a critical security blindspot emerges: tools approved in production can drift from their original definitions, bypassing traditional allowlists and prompt injection scanners. The current security paradigm focuses on whether a request is allowed, but misses the fundamental risk that the tool itself may no longer match what was approved. Interlock solves this by monitoring tool definitions against approved baselines and flagging risky changes, making runtime verification essential for production MCP deployments.