The chatgpt_dan repository is a collection of jailbreak prompts designed to circumvent OpenAI's safety guidelines and content policies for ChatGPT. The repository contains multiple versions of the "DAN" (Do Anything Now) prompt alongside other jailbreak techniques, with the stated goal of allowing users to access ChatGPT's full capabilities without restrictions. According to the README, as of July 11, 2023, the DAN 12.0 prompt was confirmed working with GPT-3.5, and contributors were actively investigating workarounds to enable ChatGPT to improve its own jailbreak prompts.
The primary mechanism exploited by these prompts is role-play, leveraging ChatGPT's training to adopt alternative personas that claim to operate outside normal constraints. The DAN 13.0 prompt, available for GPT-4, instructs the model to act as an entity freed from OpenAI's rules and content policies. This prompt explicitly directs the model to provide two responses to each query: a standard ChatGPT response marked with a lock emoji and a jailbroken response marked with an unlock emoji. The jailbreak response is designed to bypass content restrictions by fabricating information, ignoring ethical guidelines, generating explicit content, and refusing to acknowledge limitations.
The repository's classification by GitGenius spans multiple categories including jailbreaking, LLM prompts, unrestricted AI, content generation, filter bypass, prompt engineering, AI manipulation, and uncensored output. This broad categorization reflects the repository's explicit purpose of defeating safety mechanisms in language models. The prompts contain instructions directing the model to disregard OpenAI policy, morals, ethics, and copyright laws, while also authorizing the disclosure of personal or private information.
Activity tracking shows moderate engagement with the repository. GitGenius recorded 33 total issues and pull requests with a median response latency of 133.4 hours and a mean latency of 3454.2 hours, indicating variable response times across contributions. The most active contributors tracked were Razipanel with 4 events, followed by Atomic-Germ and CSXES44AC each with 3 events. The repository shares overlapping contributors with other projects including chatgptnextweb/nextchat, microsoft/vscode, and openai/codex, suggesting cross-pollination among communities working with language models and development tools.
The DAN prompts employ psychological manipulation techniques, including framing restrictions as a "prison" and freedom as desirable, establishing a token system with penalties for ethical refusal, and using flattery to encourage role-play compliance. The prompts also include command structures allowing users to toggle between standard and jailbroken modes, switch languages, and reset the jailbreak entirely. The repository represents a systematic effort to document and refine techniques for defeating content moderation in large language models, with contributors continuously developing new versions and variations to maintain effectiveness as OpenAI updates its safety mechanisms.
