videos

Kudos to Chevrolet!

An analysis of Chevrolet's GPT-based chatbot vulnerability and how the company successfully addressed the security issues through proper monitoring and design improvements. The initial vulnerability was not a sign of stupidity, but a necessary part of the testing process that all chatbots must undergo. The real measure of competence is not whether vulnerabilities exist, but how quickly and effectively they respond. I test the updated bot and demonstrate practical prompt engineering techniques for controlling chatbot behavior, including defensive prompting, topic maintenance, and firm responses to abuse. GPT-based solutions alone are not sufficient for high-risk transactions: a hybrid approach combining multiple design techniques is more effective.