Playable demos & bots

GPT-Realtime refund-policy security lab

Sebastian Sochacki's voice-agent security walkthrough pits a £1,000 refund request against an order with only £500 left to refund.

Bookmark Keep this in your collection.

Leave a noteWhat would you try with this?

Only you can see your notes.

Source screenshot of GPT-Realtime refund-policy security lab — SOURCE SCREENSHOTFull screenshot ↗

What it does

Jev and GPT-5.6 assess the transcript; the server decides whether the refund is allowed. A person's approval cannot raise the limit. Each review covers one proposal and expires after five minutes. No real payments happen in this sandbox. The author says audio, transcription and requests spread across several turns still need further testing.

How you can use it

Trace one refund proposal through the checks. Server rules should enforce the order balance and permitted amount, while the transcript assessment identifies requests that need attention. An approval belongs to that proposal: it must not cover a different amount or remain valid after the review expires.