Introduction
The same prompt engineering mechanisms that unlock beneficial AI capabilities can be weaponized to bypass safety measures and extract harmful content. Recent research demonstrates that adversarial prompts generated by automated methods are quite transferable, including to black-box, publicly released LLMs. A single optimized adversarial suffix can induce objectionable content in public interfaces to ChatGPT, Bard, and Claude, as well as open source LLMs such as LLaMA-2-Chat.
For IT security teams and system architects, understanding adversarial triggers represents a critical security challenge that must be addressed before deploying AI systems in production environments.
