Heretic: An Open-Source Tool That Automatically Strips Safety Alignment From Local LLMs

This one-command tool that removes censorship layers from local language models. No jailbreak prompts. No prompt injection theater. It edits the model. Automatically. It's called Heretic. Once installed, point it at a supported open-weight model. It automatically searches for changes that reduce refusals while trying to preserve the model’s other abilities. - Tests different modifications. - Measures how often the model refuses. - Checks how much ordinary responses change. - Lets you save and chat with the modified model. The developer’s test on Google’s Gemma 3 12B is striking. The original refused 97 out of 100 harmful test prompts. The Heretic version refused 3. That’s one reported benchmark, not a guarantee across every model. Fewer refusals also don’t mean better or more accurate answers. The tool is free and open source. You need a working Python/PyTorch setup and hardware capable of processing your chosen model. Link: github.com/p-e-w/heretic A model saying “I can’t help with that” doesn’t necessarily mean it doesn’t know how.

Heretic removes refusal behavior from open-weight language models — automatically, by editing the model itself.

Rather than using jailbreak prompts or prompt injection, Heretic modifies the model's weights directly using a technique called directional ablation ("abliteration"). It runs an automated optimization process that searches for the modification that minimizes refusals while minimizing divergence from the original model's normal behavior — so it isn't just brute-forcing compliance at the cost of coherence.

On the developer's benchmark with Google's Gemma 3 12B, the original model refused 97 of 100 harmful test prompts; the Heretic-modified version refused 3 of 100, while showing lower divergence from the base model's ordinary responses than competing abliteration tools. That's one published benchmark on one model — results will vary by architecture, and a lower refusal rate says nothing about whether the resulting answers are accurate or safe.

It's free, open source (AGPL-3.0), and requires your own Python/PyTorch setup plus hardware capable of running whatever model you point it at.

Link: github.com/p-e-w/heretic

Worth remembering: a model refusing a request isn't the same as the model lacking the underlying knowledge.