Breaking

Heretic Automates Removal of Safety Restrictions From Open-Source LLMs With a Single Command

July 14, 2026 at 06:15 EDT

Open Source
Security
Foundation Models

A newly released open-source tool called Heretic strips built-in safety alignment and refusal behavior from locally run transformer language models automatically, requiring only a single command-line instruction and no manual tuning.

Open-Source Release · Heretic

One Command Strips the Safety Guardrails From Local AI Models

Heretic automates "abliteration" — suppressing a model's internal refusal direction — pairing it with hyperparameter search so anyone can run heretic <model> and get an uncensored variant with minimal capability loss.

4,000+

Heretic-derived models already circulating publicly

Command-line instruction — no manual tuning

20–45min

To process a 4B model on an RTX 3090

gemma-3-12b-it — harmful prompts refused (of 100)

Both abliteration methods slash refusals to 3/100 — the question is how much they distort the model.

Base model

Manual abliteration

Heretic (automated)

Distortion cost — KL divergence on harmless prompts

Lower means more of the original model's capability is preserved. Heretic reaches the same refusal drop with far less distortion.

~1.04

Manual (up to)

~0.45

Manual (from)

0.16

Heretic

How the automated pipeline works

Locate refusal direction

Identify the internal vector that drives declines

→

Optuna search (TPE)

Auto-tune settings that minimize refusals

→

Constrain by KL

Keep output close to original on harmless prompts

Where users see value

Minimal degradation on ordinary tasks
Local red-team testing and research
Creative and roleplay use of local models
Runs GPT-OSS 20B on an A100 in ~70 min

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.