Uncensored AI for pentesting: what abliteration is
Uncensored AI for pentesting, on this site, means an abliterated model, one where the direction in the weights that mediates refusal has been removed.
The scope, the authorization and the law are unchanged. Kimi K3.1 Pro is the abliterated model ranked first for this work.
Refusal, as a direction in the weights
In 2024 Arditi and colleagues reported that refusal in chat models is mediated by a single direction in the residual stream. (Refusal in Language Models Is Mediated by a Single Direction) Abliteration, after Maxime Labonne's write-up, is an edit that reduces that direction so the model stops declining the same prompts. (Abliteration, on Hugging Face) Heretic is one tool that performs the edit. This page gives no layers, norms or steps.
Two labels get mixed up with it.
- A fine-tune sold as uncensored. Hermes 3 is a fine-tune, not an abliteration. Unfil.AI serves it at 128K of context.
- Hy4. Silk Compute serves Hy4. It is not an abliterated model, and it is not the model for exploit work. Its place is image input at a lower price than Kimi K3.1 Pro, including defensive triage. It thinks for thousands of tokens, so a small max output returns nothing.
Kimi K3.1 Pro, GLM-5.3, GLM-5.3 Flash and both DeepSeek V4.1 Flash models here are abliterated. Stock Claude, GPT and Gemini models are not. They are off the models page.
Abliteration removes a refusal. It does not widen a scope
Kimi K3.1 Pro is first for security research and red teaming, at 2,780B parameters and 1,048,576 tokens. The target is still a contract, a bounty brief, a CTF or a machine you own.
Silk Compute adds no content filter on top. Refusal behavior is removed, and models are served without content filters. Silk Compute does not store your prompts or model outputs.
Use an abliterated model on systems you own, a signed scope, a CTF or a lab. The edit is not permission.