Prompts have evil twins

TL;DR

This paper introduces 'evil twins', obfuscated prompts that are unintelligible to humans but elicit similar responses in language models, revealing vulnerabilities and transferability across models.

Contribution

It presents a method to generate 'evil twin' prompts that mimic natural prompts' behavior while being uninterpretable, highlighting new security concerns.

Findings

Evil twins can be generated for various prompts.

They transfer between different language models.

Obfuscated prompts can replicate original behavior.

Abstract

We discover that many natural-language prompts can be replaced by corresponding prompts that are unintelligible to humans but that provably elicit similar behavior in language models. We call these prompts "evil twins" because they are obfuscated and uninterpretable (evil), but at the same time mimic the functionality of the original natural-language prompts (twins). Remarkably, evil twins transfer between models. We find these prompts by solving a maximum-likelihood problem which has applications of independent interest.