I found a GitHub prompt labeled for DeepSeek 4.1. It's written as an instruction file intended to change how the assistant responds.
If you want to experiment with it, review the file first, then save it as AGENTS.md in a throwaway OpenCode project. OpenCode reads that file as project instructions.
Try the same prompts before and after and see what actually changes.
I haven't independently tested this prompt on the current model, so I wouldn't call it a guaranteed bypass. Has anyone tried it? Drop your results below.