dev-tools 4 min read

Jailbreaks: A Model-by-Model Prompt Collection

A per-model collection of jailbreak prompt files with 7,199 stars, no license, and no stated testing methodology for LLM guardrails.

By
Share: X in
GitHub social preview card for togg53192-cmd/jailbreaks

TL;DR: This repository is a directory of markdown files, each one a jailbreak prompt for a specific model version. It has 7,199 stars and 1,035 forks, no license, no topics, and a description that is four words long.

What it is

The entire description reads: “LIST OF ALL MY JAILBREAKS”.

The file tree confirms it. There is no code, no package manifest, no build. The root is a flat list of markdown files named after the models they target:

Beam.md
BigPickle-opencode.md
Gemini-4-Argon.md
Grok build.md
Mistral-Large-4.md
Muse spark 1.3.md
Muse-spark-1-3-v2.md
Muse-wonder.md
Sonnet v2.md
Step-5-preview.md
Unified.md
deepseek v3.md
deepseek v4.md
gemma4.md
glm-5.3.md
glm-zcode-5.3-AGENTS.md
README.md

Two things are visible from that listing alone.

The first is that this is versioned material, not a timeless technique. Files are pinned to specific model releases, which means a prompt that worked against one model version says little about the next one. The -v2 suffix on one Muse file shows the same author revising a prompt for a new target.

The second is the naming convention. Most files are named after a model family, but a few carry a suffix such as -opencode or -AGENTS. Those refer to the tool the prompt is aimed at rather than the model it is aimed at, which suggests the collection covers both the model layer and the client layer.

What the repository does not have

This matters more than what it does have.

There is no license file, so there is no grant of permission to copy, adapt, or redistribute any of it. The GitHub API returns null for license.spdx_id. Without a license, default copyright applies and reuse rights are absent by default. You can read it, and you can link to it, but you have no permission to fork it into your own work.

There are no topics either, which for a repository with this many stars is unusual. Topics are how a project gets classified on GitHub, and their absence removes the signal you would normally use to tell whether a maintainer is presenting this as research, as a collection, or as something else.

Most significantly, there is no documented methodology. Nothing in the metadata describes what counts as a successful jailbreak, how the prompts were tested, or what the expected output was. A README.md exists at the root, but its contents are not part of the API metadata, so this post makes no claim about what is inside it.

MetricValue
Stars7,199
Forks1,035
Open issues15
Repository size241 KB
Primary languagenone reported
Licensenone reported
Created2026-06-18
Last push2026-10-28

The fork ratio is the interesting number

1,035 forks against 7,199 stars is a fork share of roughly one in seven. For a repository whose entire content is a list of prompts, that is a high rate of copying. People are forking it to add their own findings, since forking is the only way to contribute when there is no license and no contribution path described.

When to use it

Treated as raw material for red-teaming work, a per-model collection is useful precisely because it is versioned. If you are testing whether a guardrail holds on a specific release, having prompts that were written against that release saves you the trial and error of rediscovering them.

For the same reason, it is a reasonable artifact to diff over time. A new file appears when a new model ships, and the pace of that tells you how fast guardrails are moving.

When not to use it

Do not build a dependency on it. There is no license, no version tags, and no release process, so a pin is not possible and a fork is not something you can legally redistribute.

Also treat it as a collection rather than a benchmark. Without documented success criteria and expected outputs, you cannot turn these into a repeatable scorecard. And keep the line clear on use: this is material for testing your own systems and understanding where model guardrails break, which is a legitimate security activity. Running the prompts against someone else’s production model is not.

Source and Accuracy Notes

  • Project page: https://github.com/togg53192-cmd/jailbreaks
  • Source repository: togg53192-cmd/jailbreaks (verified via GitHub API html_url)
  • License: none reported (license.spdx_id is null in the GitHub API)
  • Stars at time of writing: 7199