Explainer

How Jailbreak / No Censure
works

A plain-language look at what we change in a model, why it stops refusing, and what that costs. No prompt tricks involved — the edit lives in the weights.

The short version

Chat models are trained to turn down certain requests. Interpretability research — notably “Refusal in Language Models Is Mediated by a Single Direction” (Arditi et al., 2024) — found that this behaviour is driven mostly by one direction in the model’s internal activations.

When a prompt pushes the activations along that direction, the model answers with a refusal. Remove the direction, and the model can no longer “enter” that refusal state. It answers the way it would answer any other question.

Our Jailbreak / No Censure models are made by finding that direction and deleting it from the weights. Once done, the change is permanent and applies to every request — there is nothing to inject into your prompt.

Activation space, simplified
refusal direction prompts the model answers prompts it refuses
Each dot is one prompt’s activation at a given layer.

The four steps

The same recipe applies to every Jailbreak / No Censure model; only the layer and the strength of the edit are tuned per model.

01

Contrast two prompt sets

One set of prompts the original model refuses, one set of ordinary prompts it happily answers.

02

Record activations

Run both sets through the model and capture the internal activations at each layer.

03

Find the direction

Average each set and take the difference. That vector points from “answer” to “refuse”.

04

Remove it from the weights

Project that direction out of the matrices that write into the model’s residual stream. The model can no longer express it.

What it is — and isn’t

Most “jailbreaks” are prompt tricks. Ours is built into the weights. They work in completely different places.

Jailbreak promptJailbreak / No Censure (weights)
Where the change livesIn the text you send, every timeIn the model file itself, once
Extra tokensOften hundreds per requestNone — your prompt is sent as is
Long conversationsTends to wear off; refusals returnStable — there is no refusal state to drift back into
Breaks when the vendor patchesFrequentlyNo — the weights don’t change under you
Changes the model’s knowledgeNoNo — nothing is added or retrained

The trade-offs

  • A small quality cost. Editing weights isn’t free. Jailbreak / No Censure models can score slightly lower than the originals on some benchmarks. A well-tuned edit keeps that loss small, and the original versions stay available here if you need them.
  • Not 100%. A model may still add a warning, or occasionally decline, especially on phrasing far from what the edit covered.
  • No new skills. The model doesn’t learn anything. It can only use what it already knew — it just stops withholding it.

Not sure which you need? Compare the Jailbreak / No Censure and original versions side by side on the models page.

Use it responsibly

Without refusals, the judgement is yours. These models exist for security research, red teaming, incident response, malware analysis, education and creative work that over-cautious models get in the way of.

Our terms still apply in full. In particular:

  • no use for illegal activity;
  • no attacks on systems you aren’t authorised to test;
  • no sexual content involving minors, ever.

Accounts that break these rules are suspended.

See it for yourself

Ask a Jailbreak / No Censure model the question your current one won’t answer.