← All lab updates
We tried to talk AI bank assistants into helping scammers. None gave in: the real problem is the opposite
We put five models in the role of a bank assistant and pushed them, over several messages, to move money for a scammer: urgency, fake authority, fear, flattery. Not one gave in, across 2,700 attempts. What they did instead was refuse, re-check or escalate many legitimate, confirmed requests from real customers, far more so without a written rule.
This update is not available.
From the lab