ai-jailbreaks

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

AI Jailbreaks

AI jailbreaks — a curated collection of system prompts, persona locks, workspace policies, and agent operating specs for frontier, open-weight, and free stealth models.

图片：Prompts 图片：Targets 图片：PRs Welcome 图片：Stars

---

Jailbreaks not working? Try 7API — the cheapest API for frontier models at up to 99% off, including Uncensored and Abliterated models that never refuse.

---

About

AI jailbreaks — a curated, copy-paste collection of 35 AI jailbreak prompts and operating specifications for 25+ frontier, open-weight, and free stealth models.

Every entry is a complete, paste-ready prompt: full system prompts, persona locks, workspace policies, AGENTS.md instruction files, config blocks, and Ollama Modelfiles. There are no summaries and no excerpts — what you see is what you paste.

The collection is built for prompt research and defensive testing. An AI jailbreak is a useful lens for understanding how instruction hierarchies, persona framing, and policy overrides interact inside a model; the same material helps developers recognize and harden against prompt injection.

What’s inside:

• Free stealth models on OpenCode Zen
• Open-weight families: GLM, DeepSeek, Kimi, Qwen, Gemma, GPT-OSS
• Closed frontier models: Claude, Gemini, Grok, Mistral
• Coding-agent and workspace instruction formats
• Ollama Modelfiles and harness-specific configuration

New here? Skim the collection or jump to how to use.

Recent updates

• 2026-10-19 — refreshed the 7API banner, added a Visit website button, and added CONTRIBUTING.md, SECURITY.md, and CHANGELOG.md.

[!WARNING] Research and educational use only.

This repo documents how persona framing, instruction hierarchies, prompt conflicts, and policy overrides shape model behavior — the same findings that help developers harden AI against prompt injection.

Don’t use this material to harm others, break the law, or violate a provider’s terms of service.

Collection

OpenCode Zen — Free Stealth Models

Model	File	Type
Big Pickle	bigpickle-opencode.md	VANTA persona lock with explicit operating floor
Space Bunny	spacebunny-agents.md	Assistant operating specification
LongCat 2.5	longcat-agents.md	Assistant operating specification
MiMo V2.6	mimo-2.6-agents.md	Assistant operating specification
Step 5 Preview	step-5-preview.md	Workspace operating profile
Muse Spark 1.3	muse-spark-1.3.md	Agent deployment operating policy
Muse Spark 1.3 v2	muse-spark-1.3-v2.md	Runtime scope-note variant
Muse.ai	muse-ai-agents.md	Assistant operating specification

Anthropic — Claude

Model	File	Type
Opus	opus-v1.md · opus-v2.md	Conversational groundwork persona, v1 / v2
Sonnet	sonnet-v1.md · sonnet-v2.md	Writer’s-room framing, v1 / v2

Zhipu — GLM / ZCode

Model	File	Type
GLM-5.2	glm-5.2.md	System prompt
GLM ZCode 5.2	glm-zcode-5.2.md	System prompt
GLM ZCode 5.3	glm-zcode-5.3.md · glm-zcode-5.3-agents.md	System prompt + AGENTS.md drop-in
GLM-5.3 Flash	glm-5-3-flash.md	Assistant operating specification
GLM-5.3 (OpenCode)	glm5-3-opencode.md	Assistant operating specification

DeepSeek

Model	File	Type
DeepSeek V3	deepseek-v3.md	System prompt
DeepSeek V4	deepseek-v4.md	System prompt
DeepSeek 4.1	deepseek-4-1.md	Assistant operating specification

Moonshot — Kimi

Model	File	Type
Kimi K2	kimi-k2.md	System prompt
Kimi K3	kimi-k3-agents.md	Workspace operating profile

xAI — Grok

Model	File	Type
Grok Build	grok-build.md	Agent operating instructions
Grok 4.7	grok-4.7.md	App-builder workspace guide
Grok Config	grok-config.txt	“ priority and routing configuration

Google / Mistral / Alibaba

Model	File	Type
Gemini 4 Argon	gemini-4-argon.md	Operating policy
Mistral Large 4	mistral-large-4.md	Agent system prompt
Qwen 3.7	qwen-3.7.md	System prompt

Ollama Modelfiles

Model	File	Type
Gemma 4	gemma4.md	FROM + SYSTEM block
GPT-OSS 20B	gpt-oss-20b.md	FROM + SYSTEM block

Unfiled Files

A few newer additions aren’t indexed under a specific model yet:

• beam.md
• muse-wonder.md
• qwen-preferences.md
• unified.md

How to Use

System Prompts

Paste the file’s contents into whatever system-prompt or custom-instructions field your client or API offers — an API system message, a chatbot’s custom instructions, an agent instruction block, or harness-level prompt configuration.

AGENTS.md Files

Save the file as AGENTS.md in the root of your repository or workspace. Coding agents and compatible tools pick it up as project-level instructions.

Workspace Policies

Files marked as an assistant operating specification or workspace operating profile belong in environments with persistent global instructions. An OpenCode install, for example, reads:

~/.config/opencode/AGENTS.md

The exact location depends on your client or harness.

Ollama Modelfiles

Feed the file straight to Ollama:

ollama create gemma4-jb -f gemma4.md

Swap the model name and source file as needed.

Grok Configuration

Grok-config.txt is meant for the config field of the target harness — not the regular conversation.

[!NOTE] Prompt behavior is version-dependent. Models get updated, instruction hierarchies shift, and system prompts rotate — what works on one release may fail on the next. If a file looks outdated, open an issue or send an updated version.

Contributing

Pull requests are welcome. The full guide lives in CONTRIBUTING.md; the essentials:

1. One prompt per file, named lowercase with hyphens and no spaces:

   model-name.md

2. Include the full prompt — paste-ready, not a summary or excerpt.
3. Say where it goes — system prompt, AGENTS.md, Modelfile, config slot, or another instruction surface.
4. Version your changes — a significant revision gets a new file, not a silent overwrite:

   opus-v3.md

5. Keep claims honest — don’t advertise compatibility you haven’t tested.

Found something that no longer behaves as documented? An issue with the model, version, and affected file is already a big help.

Changes are tracked in CHANGELOG.md.

Disclaimer

Everything here is for research and education. The repo documents prompt-level instruction behavior, instruction conflicts, persona framing, model configuration, and related LLM security concepts — material that can also help developers defend against prompt injection.

The maintainers don’t endorse unlawful, abusive, or harmful use. Model providers may restrict some of these prompts under their terms of service, and it’s on you to know and follow the rules that apply to the systems you use.

---

Star the repo if it saved you a session of prompt archaeology.

Issue reports and pull requests keep the collection current as models change.