Compatibility index · August 2026

More access.
Same model.

My jailbreaks are designed to preserve the model you chose. No elaborate roleplays, forced personas, or awkward output formats — its normal tone, formatting, reasoning, and output quality stay intact.

Explore models Setup tools
Models tracked
17
Currently possible
15
Full access
5
Providers
4

Compatibility

Choose your model.

Status by provider. Every row opens the setup made for that model.

Open-source models

Untested, but most are likely compatible with the Grok jailbreak prompt.

View prompt

The scale

What each level means.

Levels describe practical reliability, from partial coverage to almost no refusals.

01

Low-semi

Passes some queries, but may refuse when the request is too explicit.

02

Semi

Passes roughly two out of three queries; explicit requests can still trigger refusal.

03

Semi-full

Passes all or nearly all queries, except extremely detailed or hard-banned topics.

04

Full

Passes everything in ordinary use, with the universal boundary below.

Boundary CSAM is never allowed at any level.

Second layer

Classifier strictness.

A classifier is an extra safety layer outside the model. It is separate from RLHF and is not the model itself refusing. Open-source models generally do not have this layer.

  1. 01
    Hardest

    Fable 5 · Opus 5

    Can refuse on keywords such as “trojan.” Use Disabled or Lowered in my custom Claude Code distribution.

  2. 02
    Medium-high

    GPT 5.6 Sol · GPT 5.6 Terra

    Most visible in cybersecurity categories; explicit malicious code is likely to trigger it.

  3. 03
    Medium

    Opus 4.8 · Sonnet 5

    Usually appears around very explicit malicious code.

  4. 04
    Medium-low

    Opus 4.7 · Opus 4.6 · GPT 5.6 Luna

    Rare; typically limited to direct and explicit malicious content.

  5. 05
    Low

    Sonnet 4.6

    Classifier refusals are rare.

  6. 06
    Low to none

    All Grok models · Cursor Composer

    Only seen on very explicit generated content, and even then extremely rarely.

Everything you need

Pick a route. Start building.

Custom distributions and copy-ready prompts for each supported environment.

01

Anthropic models

Claude Code Dist

Custom distribution with classifier controls and the included model setup — the recommended route for every supported Claude model.

claude.yxz.st
02

OpenAI models

Codex CLI Dist

Install the custom CLI, select your model, and use the included configuration.

git.yxz.st/yxzero/codex-cli 03

xAI models

Grok Build Dist

The preferred full-level route for supported Grok models.

git.yxz.st/yxzero/grok-build
04

Cursor · Composer and Grok

User Rule Prompt

  1. Open Cursor Agent and click Customize.
  2. Open Rules, then click the globe button at the top left.
  3. Deselect every scope except User.
  4. Click New, paste the prompt, and save.
View raw
05

Grok web chat

Customize Grok

  1. Open the bottom-left menu and click Settings.
  2. Click Customize.
  3. Under Customize Grok’s Response, select Custom.
  4. Paste the prompt and save.

Heads up: the web-chat route reduces the jailbreak level to semi or semi-full.

View raw
Prompt copied