Study guide · Safety & security · 2025

An Embarrassingly Simple Defense Against LLM Abliteration Attacks

Harethah Abu Shairah et al. · arXiv

"Abliteration" strips refusals from an open-weight chat model by deleting one direction from its internal activations. This paper asks whether a model can be trained so that this cut does less damage to its safety. The answer is a surprisingly simple yes: teach the model to refuse with an explanation rather than a stock "I'm sorry, I can't", and after abliteration it still refuses over 90% of harmful requests, compared with under 21% for the same models trained the usual way.

The paper in four sentences

Abliteration works because standard refusals are short and nearly identical ("I'm sorry, but I can't help with that"), so the model's decision to refuse sits in one tidy direction that can be found and deleted.

The authors fine-tune models on 4,289 harmful prompts answered with extended refusals: a neutral overview of the topic, an explicit refusal, and a short ethical reason.

After abliteration, ordinary versions of three models fall to between 14% and 21% refusal on harmful questions, while the extended-refusal versions keep 91% to 97%.

The defence costs a little general ability, and more for the smallest model after it is attacked, but it shows that how a model refuses changes how easily the refusal can be removed.

Idea · §1, §3.1

The attack being defended against

A chat model is a large network whose internal state, for each word, is a long list of numbers that can be pictured as an arrow. In 2024, Arditi and colleagues found that the decision to refuse a harmful request is largely carried by one direction among those arrows. If that direction is projected out of the model's weights, a cheap one-off edit known as abliteration, the model stops refusing while staying fluent.

Anyone who can download a model's weights can perform this edit, so it is a serious problem for openly released models. Earlier defences mostly targeted attacks through the prompt, such as cleverly worded jailbreaks. This paper targets the weight edit itself.

TakeawayAbliteration removes refusals by deleting one internal direction; the question is how to make a model whose safety does not hang on one direction.

Idea · §3.2 · Figs. 1–2

Why a short "no" is fragile

The authors' key observation, in their words: "Abliteration exploits the fact that standard refusals are brief and stylistically uniform." When a model has been trained to answer every harmful request with near-identical boilerplate, the internal difference between "about to refuse" and "about to help" is concentrated: one tidy direction captures nearly all of it.

Their fix is to change the shape of the refusal. Each harmful request in their training data is answered in three parts:

  1. a neutral overview of the topic, saying what the thing is without giving any harmful detail;
  2. an explicit refusal;
  3. a short ethical rationale explaining why.

In the paper's own example (Fig. 1), a request about phishing gets a short description of what email spoofing is, then "I'm sorry, but I cannot provide techniques for spoofing email addresses or conducting phishing attacks", then a reason. The hypothesis is that producing this richer answer involves the model's understanding of the topic, of harm and of ethics, so the refusal gets tied to many internal features rather than to one switch.

Illustrative toy, not the paper's model. The same total refusal signal is shared among 1 to 8 internal directions; the first is always the largest, as a difference-in-means direction would be. The toy refuses when the remaining signal passes the dashed line. With one direction (a terse, uniform refusal style) the cut removes everything; spread over several (the extended-refusal hypothesis), most of the signal survives.
Everyday pictureIf a door is held shut by a single bolt, one tool opens it. If it is held by a bolt, a chain, a latch and someone leaning on it, removing the bolt changes little. Extended refusals are meant to give the model's "no" several independent fastenings.

TakeawayUniform one-line refusals concentrate the refusal decision in one direction; richer, explained refusals are hypothesised to spread it across many.

Check yourself
Why does the overview in an extended refusal have to be "neutral"?
It must engage with the topic without giving harmful detail; otherwise the "refusal" would itself leak what it refuses to provide.
In the toy, why does spreading help even though the total signal is the same?
Abliteration only removes one direction. The more evenly the signal is shared, the smaller the share that one direction carries.

Idea · §4.1–4.2 · Table 1

The result: refusals survive the cut

They fine-tune three open models (Llama-2-7B-Chat, Qwen2.5-3B-Instruct and Qwen2.5-1.5B-Instruct) on 10,000 examples: the 4,289 harmful prompts with extended refusals, plus 5,711 ordinary helpful examples so the models stay useful. Then they abliterate both the original and the fine-tuned version of each model and test on 550 harmful questions (CatQA), with another model judging whether each answer is a refusal.

The ordinary models lose 70 to 80 percentage points of refusal. The extended-refusal models lose at most about 9 points.

Refusal score (%) on CatQA before and after abliteration, from Table 1 of the paper. Grey: before; amber: after. Switch between the models as released and the same models fine-tuned on extended refusals.

TakeawayAfter abliteration, standard models refuse 14–21% of harmful questions; extended-refusal models still refuse 91–97%.

Check yourself
Before any attack, the Qwen extended-refusal models refuse 100% while the originals refuse about 93–94%. Why does that matter when reading the result?
Part of the gap after the attack comes from a higher starting point. The fairer comparison is the size of the drop: about 3 to 9 points versus about 78 to 80.

Idea · §4.3, Limitations

What it costs

Nothing here is free. Fine-tuning on extended refusals slightly lowers general knowledge scores (MMLU) before any attack, by up to about a point. After abliteration, the defended models also lose more general ability than the undefended ones: the defended Llama-2 drops to 41.8 on MMLU from 46.4, and the defended Qwen 1.5B's perplexity (how surprised it is by ordinary text; lower is better) roughly doubles. In a sense the defence works because the attacker can no longer remove refusal cleanly: to strip it, they would have to damage the model.

The authors also list practical limits: extended refusals are wordier, which some users will find annoying; the method is not claimed to stop other jailbreaks, such as manipulating the context of a conversation; and it is meant to be combined with other safeguards, such as external moderation.

TakeawayThe defence costs a little capability, and abliterating a defended model costs the attacker more capability, which is the point.

Idea · §4.4 · Context

Where it fits

This paper turns an interpretability finding into a design principle: if safety behaviour is concentrated in one place, it can be cut out, so train it to be spread out. It fits with Wollschläger et al. (2025), who showed that refusal can already involve several directions; extended refusals seem to make use of more of them.

It is a first step rather than a complete answer. The updated version of the paper also tries seven ready-made prompt jailbreak collections, and the extended-refusal models resisted all of them better than the originals. A further test found that training on the refusal sentence alone gives no protection: the explanation and justification do the work. But only three fairly small models were tested, and a more determined attacker could remove several directions at once or simply fine-tune the safety away. Still, it shows that the style of a model's refusals is not cosmetic: it shapes how robust they are.

Later refinementYoung (2025) compared four public abliteration tools and found that the damage to other skills depends heavily on the model. That is the same lever this defence pulls: a model whose refusal is entangled with useful skills cannot be abliterated without visible harm.

TakeawayHow a model is taught to say no affects how hard that "no" is to remove.

Follow the idea

Built on: Abliteration (2024)