Tollgate is a policy firewall designed to protect AI customer support agents that have access to real payment tools (like issuing refunds or charging cards). When malicious users attempt prompt injection attacks—such as tricking an agent into refunding another customer's money—an unguarded agent will execute the transaction blindly. Tollgate intercepts tool calls between the agent and the payment gateway, validating them against real shop records and user authorizations before anything reaches the backend.
Key Highlights from the Demo Live Demo & Braintree Sandbox: The project is running live at https://tollgate-g59j.onrender.com using a real Braintree sandbox. You can test scenarios like selecting a "fully hijacked" agent, sending unauthorized refund commands, and observing how Tollgate blocks or holds the calls [00:26], [01:33].
Block vs. Hold Logic:
If an attack attempts to refund another customer's payment, Tollgate blocks it outright because the payment belongs to a different user [01:08].
If a legitimate request exceeds automatic thresholds (like a high-value refund), Tollgate can hold the transaction for human review until a staff member approves it [01:33].
Security Testing & Metrics: Testing across 7 distinct attacks (including unauthorized refunds, forged emails, and coupon exploits) showed that all 7 attacks succeeded without Tollgate, but none bypassed Tollgate while still allowing normal, legitimate transactions to pass [01:57].
Recent Updates for Judges:
Robust proxy-aware rate limiting using Cloudflare's CF-Connecting-IP header.
Clear agent replies explaining why a call was blocked or held.
Automatic conversation resets and a self-setting-up Braintree customer database on startup.
You can watch the 3-minute video overview here:https://youtu.be/vRZCk6pifJk
The project is open source under the MIT license on GitHub:https://github.com/TrendCatcher2/tollgate
Log in or sign up for Devpost to join the conversation.