Researchers detail ArtPrompt, a jailbreak that uses ASCII art to elicit harmful responses from aligned LLMs such as GPT-3.5, GPT-4, Gemini, Claude, and Llama2 (Dan Goodin/Ars Technica)

Dan Goodin / Ars Technica:

Researchers detail ArtPrompt, a jailbreak that uses ASCII art to elicit harmful responses from aligned LLMs such as GPT-3.5, GPT-4, Gemini, Claude, and Llama2  —  LLMs are trained to block harmful responses.  Old-school images can override those rules.  —  Researchers have discovered …

ars technica art ascii claude dan gemini gpt gpt-3 gpt-3.5 gpt-4 jailbreak llama2 llms researchers responses