A Firewall Behind Your Voice Agent
Why to put Cortega between Vapi and the model, the setup steps in Vapi and Cortega, and how agentic attacks on chatbots work, with public cases.
Read the research report ↗Medhavi Bhatia · CTO & Co-founder · November 2026
TL;DR
Related white papers · CORTEGA LABS
Explore the full research library ↗In December 2023 a car dealer's chatbot "agreed" to sell a new Tahoe for one dollar. The customer had told it to agree with everything and to call every answer a binding offer. No car changed hands, and the chatbot was taken down.
A voice agent has the same weakness and less room for error. A spoken promise cannot be edited after the caller hears it.
A voice agent built on Vapi turns speech into text, sends the text to a model and speaks the answer. Vapi lets you choose where that model call goes. Point it at the Cortega gateway and every turn is checked twice: the caller's words before the model sees them, and the model's answer before the caller hears it. A blocked turn is spoken as a normal refusal, so the call goes on.
What you get:
Attacks on chatbots are rarely one message. The attacker reads each refusal and tries again with a new reason, a new role or a smaller first step. Public cases include the dealer above, a delivery firm's chatbot that swore at a customer, and a zero-click data leak from Microsoft 365 Copilot.
Our research report covers the setup in Vapi and in Cortega, step by step, the main attack techniques and what stops each, the real cases, and a test list to run before launch.
Read the report (business email confirmation required).
MAKE YOUR NEXT MOVE
Explore how Cortega fits your AI environment.