upstreamism, Researchers asked an AI chatbot to act as a research assistant, then instructed it to develop prompts that could 'jailbreak' other chatbots so that they would produce instructions for making meth, laundering money, and building bombs. The approach had a 42.5% success rate against GPT-4. https://www.scientificamerican.com/article/jailbroken-ai-chatbots-can-jailbreak-other-chatbots/
Add comment