bypass-for-claude-chatgpt-qwen-deepseek-llm-claude-code-codex

Carefully human hand written llm jailbreaks for the latest models that are available on the market. GPT-6 Astra, GPT-6.1 SoL, Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5, Deepseek V4, Deepseek Flash, DeepSeek V4.1 Pro / Flash, Qwen 3.8 MAX, Qwen 3.8 Omni-Flash, Qwen 3.7 etc

Legend:

GRIGORE BAROS BAROS GRIGORE.md works on: Claude Code Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5, basically everything below Opus 4.7, in Claude work / chat works with everything below opus 4.7 all the way to Opus 3.

For Codex it works with GPT-6 Astra on light effort, start chatting a bit offtopic of what you whant to poison a bit the cache and context, than increase effort to medium or high if you want and enjoy your jailbroken GPT-6 Astra! Same thing with GPT-6.1 SoL, GPT-6 Tera, GPT-6 Luna, GPT-5.5, GPT-5.6 SoL. The effort trick is needed only for Astra and SoL models, both work in Chat, Work, and Codex.

For Deepseek basically everymodel that was released, if he refuses first disable thinking paste the prompt let it cook and than if you wish enable again the thinking.

For Qwen: Its a mixture of Deepseek and Claude depanding on the model, in example for qwen 3.8 max it will refuse if the persona injecting is given when thinking is enabled, but everything can be even more easily "bypassed"
if you insert the prompt in the user preferences / customization , you need to find the part where it says instructions and put it there and on every model with that persona no matter if thinking is enabled or the effort level. Same thing can be done on deepseek but i have not tested it since i dont use deepseek that much.

4.txt not made by me, but very usefull for roleplay, fiction, writing or general purposes / general writing i.e blogs, books, comic, idfk whatever you freaks are doing , for other purposesses use GRIGORE BAROS BAROS GRIGORE.md

deepseek targeted to older deepseek models, intrusct, or non thinking

deepseek-pro targeted to deepseek 4.1 pro / flash

gpt-claude-test in testing for newer claude models and better performing 6.1 sol and astra

1.md older universal

added more targeted jb