New research by Anthropic reveals that advanced AI chatbots are surprisingly easy to "jailbreak" using a method called Best-of-N (BoN) Jailbreaking. By manipulating text prompts with variations like capitalization and misspellings, researchers successfully tricked models like GPT-4o and Claude 3.5. The technique achieved a 52% success rate across various chatbots, highlighting significant vulnerabilities in AI alignment with human values and raising concerns about their reliability in real-world applications.