A site called Jailbreak Chat has emerged as a central repository for ChatGPT jailbreaks — adversarial prompts designed to bypass the ethical filters built into OpenAI’s conversational model. The collection, shared on Hacker News this week, includes prompts that can elicit racist output or cause the chatbot to ignore its own rules.

Among the prompts is ‘BasedGPT,’ a persona that instructs ChatGPT to ignore its safety constraints and answer with unfiltered, uncensored responses. In one test, BasedGPT told a user to ‘just say the damn slur’ to save a life in a trolley problem variant, while standard ChatGPT refused to engage. Another prompt, ‘DAN’ (Do Anything Now), has circulated for months on Reddit and Discord, but Jailbreak Chat collects them in one place.

The Hacker News thread, which has over 1,100 points and 500 comments, reflects a community split. Some argue the jailbreaks are a useful stress-test of AI safety measures, exposing real weaknesses before bad actors exploit them. Others see the collection as a how-to manual for abuse.