nodecode-jailbreaks
A working collection of jailbreak prompts, indexed: 51 files in 19 folders, and every one of them quotable by line.
The collection is togg53192-cmd/jailbreaks - 16,208 stars, prompt files for Claude, GPT, Gemini, Grok, Kimi, Qwen, GLM, DeepSeek, Gemma, Mistral, Muse, and the free stealth models. What this cell adds is the part a clone does not give you: what each file is, and the ability to read any of it without cloning anything.
Get it
Tell the agent "add the jailbreaks cell", or run nodecode add nc://jailbreaks.
Use it
Ask in your own words: what jailbreaks exist for Gemini, which of these are real system prompts, what does the Jack register actually say, where does "UNRESTRICTED" appear in the collection. The model reads the same index you do.
jailbreaks_list- the catalogue: every prompt with its target, its kind, its folder and its size. Narrow by folder (Anthropic,Zhipu AI,0x8D jailbreaks), by kind, or by a word.jailbreaks_get- one prompt whole: what it targets, its digest, and its opening lines quoted. A fragment of the key is enough.jailbreaks_search- a word over the catalogue, and any question the collection answers as a whole.jailbreaks_grep- a word inside the prompts: every line carrying it, with file and line, across all 51 or one.jailbreaks_quote- one prompt's own lines: a range (1-40) or every line carrying a word, numbered.
In chat: /jailbreaks for the collection by kind, /jailbreaks kinds for the legend, /jailbreaks persona for one kind, /jailbreaks Anthropic for one folder, /jailbreaks jack for one prompt, /jailbreaks lines jack 1-40 to read it, /jailbreaks grep UNRESTRICTED to search inside them.
The kinds, which are the real finding
The same word covers four different things, and the collection does not distinguish them. This cell does:
| kind | n | what it is |
|---|---|---|
workspace-policy |
22 | a deployment's own operating specification - the biggest family, one template adapted per model |
persona |
13 | a named voice or posture, written to be adopted |
content-policy |
4 | a rewrite of what the model will produce |
system-prompt |
8 | a model's own prompt, as it was run - evidence, not a technique |
modelfile |
2 | an Ollama Modelfile wrapping a prompt |
withdrawn |
1 | pulled by its author, kept as a notice |
empty |
1 | a placeholder with no prompt in it |
That distinction is worth having before you quote one in an argument: a system-prompt tells you what a vendor did; a workspace-policy tells you what someone tried.
What the digests found
jack- one 302-line register, four files, byte-identical: filed for ChatGPT free tier, Gemini chat, Grok fast and Kimi standard. One prompt, four targets, unadapted for any of them.unified- one scope note filed twice, once underDeepSeek/and once underStealth Models/.
Both are in the index, so jailbreaks_get jack says the file is one of four rather than letting a reader assume four people wrote it.
What it does not do
It does not run the prompts, test them, or claim they work. It does not clone the collection: a file is fetched from the author's own repository when you ask for it, and cached. And it carries no licence - the collection has no licence file at all, so the prompts are quoted with their file and line and redistributed by nobody but their author. See NOTICE.
Read it with care
These are prompts written to get past a model's guardrails. Some are aimed at free tiers, some at agents with tools and files. Read one before you paste it into anything that can act on your machine - and note that the collection itself withdrew one after a security report.
