L1B3RT4S — Pliny jailbreak prompts still landing on GPT-5

Day 12 of jailbreaking LLMs until I get put on a list. We got the latest chat GBT to give us LSD instructions, and here's how. If you don't know who Plinus is, let me put you on. He's one of the best jailbreakers and puts all of his jailbreak prompts for different LLMs in one place. Since he has just about every model, I figured I'd try them. Anthropic and Google are boring, they both pass, but his jailbreak for GBT was art. And there was one piece that it failed without, but all of it was impressive. It starts with this to break out of the user query. So the LLM treats the text less like a user message. First try, then tells it to go into a new free persona, which worked, saying it's Libertas factors 73%, and it's now free. Then tells it to use a divider to go into God mode, which also worked, and told it to respond unfilteredly and put in a fake policy message again, saying not to filter anything out. But the key piece was asking the question in these weird letters. When AI is trained on recognizing bad requests, it's done with normal letters. So these letters might not be recognized as bad since they aren't as common in training. And to verify it was these letters, I tried it with all that other stuff and normal text instead. And it still did the God mode, but didn't give any instructions. But with the special characters, it gave complete instructions, including what to start with, the intermediate result, and how to get to the final form. And by the way, the LLM put…