Prompt Injection Attacks Against GPT-3

A fascinating series of posts from Simon Willison about attacks with malicious prompts for automated responses based on machine learning — the second and third parts are linked in the sidebar.

Fascinating and troubling to consider this as a parallel to social engineering attacks on real, living people. It is not a stretch to imagine more call centre tasks being offloaded to automated systems — regrettably.1 Agents are trained to avoid divulging information like the customer’s address or partial credit card number, but too heavy reliance on prompt-based tasks might result in an uptick of these kinds of attacks.

  1. The loss of employment for millions is an obvious concern. On the other side of the phone line, there is a satisfaction difference. I have spent the past couple of weeks on the phone with various call centres, and there is a vast gulf in my level of happiness between speaking with a real person and speaking with a robot for even part of it. ↥︎