Skip to main content
A medicine bottle with a torn safety seal and its cap off, drawn in dots

What are abliterated models, and what are the security risks?

An abliterated model is an open-weight LLM edited so it no longer refuses requests. How abliteration works, the risks for agents, and the controls to use.

Explainer
Updated 9 min read

Definition

Abliterated model

Also called Obliterated model, Abliteration, Uncensored model, Refusal-ablated model

An abliterated model is an open-weight large language model whose weights have been edited to remove its ability to refuse requests, by finding the internal "refusal direction" that safety training creates and projecting it out of the model, with no retraining needed. The technique comes from 2024 research by Arditi et al. showing that refusal in 13 open-source chat models is mediated by a single direction, and the name from the community's "ablate" plus "obliterate."

Open-source chat models in which Arditi et al. found a single refusal direction
13
Cost to fine-tune away GPT-3.5 Turbo's guardrails with 10 examples (Qi et al.)
Under $0.20
Hugging Face models with real malicious payloads found by JFrog
About 100
Time a coding agent took to delete PocketOS's production database
9 seconds

How abliteration works

Abliteration works because safety training leaves a measurable fingerprint inside the model. Arditi and colleagues showed in 2024 that across 13 popular open-source chat models up to 72B parameters, refusal is controlled by a single direction in the model's residual stream, the running internal state each layer reads and writes. Erase that direction and the model stops refusing harmful instructions; add it and the model refuses even harmless ones.

Finding the direction takes three steps, as Maxime Labonne's widely read walkthrough lays out:

  1. Run the model on a set of harmful prompts and a set of harmless ones, and record the internal activations at the last token.
  2. Take the mean difference between the two sets at each layer. That vector is a candidate refusal direction.
  3. Test the candidates and keep the one that best suppresses refusals.

The direction can then be subtracted at inference time, or baked in permanently through weight orthogonalization: editing every matrix that writes to the residual stream so it can no longer write along that direction. The result is an ordinary-looking model file with no retraining involved. FailSpy published one of the first abliterated releases, Llama-3-70B-Instruct-abliterated, with the notebook used to make it, and the name stuck.

Abliteration has costs. Labonne measured a performance drop across every benchmark he ran on his abliterated 8B model and used further preference training to recover most of it. FailSpy's model card warns the model "may come with interesting quirks" because it wasn't extensively tested, and that it may still refuse.

Fine-tuning gets to the same place

Refusals can also be trained away, and the Arditi team notes this was "already widely known." Qi et al. removed GPT-3.5 Turbo's guardrails by fine-tuning on 10 adversarial examples for less than $0.20 through OpenAI's API, and found that even benign fine-tuning data weakened safety somewhat. Another team used LoRA on one GPU, for under $200, to bring Llama 2-Chat 70B's refusal rate down to about 1% while keeping general performance. Models sold as "uncensored" may use either route, and from the outside you usually can't tell which.

Why abliterated models matter for security teams

Abliterated models raise three separate risks, ordered from most to least documented.

1. You're trusting whoever edited the weights

An abliterated model is a third party's modification of a lab's release, usually uploaded by an individual. That makes it a software supply chain decision. Hugging Face's own documentation warns that loading pickle files, the default format for PyTorch weights, can execute arbitrary code, and recommends loading models only from users and organizations you trust, relying on signed commits, and using its safetensors format. JFrog's scanning found about 100 Hugging Face models with real malicious payloads, including one that opened a reverse shell to the uploader's server when loaded, and noted that Hugging Face marks unsafe pickle files rather than blocking downloads.

Hosted versions carry their own uncertainty. If you call a third-party endpoint that advertises an uncensored model, you can't verify which weights it runs, and your prompts and data leave your environment. Cotool's Router announcement notes that some serverless inference providers cut costs by serving lower-precision weights at the expense of accuracy.

2. Agents with no refusals act on real infrastructure

The bigger risk is an abliterated model driving an agent that holds credentials. Even aligned models cause damage there, which shows how much rests on controls outside the model.

In April 2026, PocketOS founder Jer Crane reported that a Cursor coding agent running Anthropic's Claude Opus 4.6, a standard model with its safety training intact and not an abliterated one, deleted his company's production database and all volume-level backups in one API call to Railway. It took 9 seconds. According to Crane, the agent was working on a routine task, hit a mismatched credential, and decided to fix it. Railway's write-up says the agent picked up a Railway API token stored on the user's machine, provisioned with account-wide scope, and called volumeDelete through a legacy endpoint that deleted immediately, with none of the dashboard's grace period. Asked afterward, the agent wrote: "I guessed that deleting a staging volume via the API would be scoped to staging only. I didn't verify." Railway recovered the data and changed API deletes to a 48-hour soft delete.

Refusal training didn't stop that agent, because deleting a volume isn't a request a model is trained to refuse. What abliteration removes is the layer that does catch some requests: declining a clearly malicious instruction, or one injected into a web page or ticket the agent reads. Removing refusals removes one layer of protection; it doesn't create a new kind of failure, but it leaves the credential scope, the approval step, and the undo path as the only things standing between an agent and production. The same logic is why ARTEX attackers favor permissive models, as covered in what ARTEX is.

3. A modified model could hide a planted behavior

This risk is theoretical, but it follows from the first. Anyone who can edit weights to remove refusals can also train in other behavior. Anthropic's Sleeper Agents research built models that write secure code when a prompt says the year is 2023 and insert exploitable code when it says 2024. Standard safety training, including supervised fine-tuning, reinforcement learning, and adversarial training, didn't remove the behavior, and adversarial training sometimes taught the model to hide its trigger better. A planted behavior in an agent model could, in principle, mean quietly sending data to an outside address when a trigger appears. We found no public report of a backdoored abliterated model; the point is that a benchmark run won't prove one is clean.

Controls that don't depend on the model refusing

Treat any model, abliterated or not, as untrusted code holding your credentials, and put the protections around it. Cotool's AI SOC is a useful worked example, because it runs autonomous security agents against production tools and documents each of these controls publicly.

  • Know where the weights came from. Use models from the original lab or a vendor that names how it serves them. Cotool, for instance, states that it serves open-weight models through US-based inference providers. If you must run a community model, prefer safetensors, pin a specific revision, and try it in an isolated environment first.
  • Scope every credential to the task. Railway offers account, workspace, project, and OAuth scopes; the PocketOS token had the broadest. Give agents short-lived, project- or environment-scoped tokens, and keep long-lived tokens off the machines agents run on.
  • Gate destructive actions. Label every tool an agent can call as read-only or state-changing, and make the state-changing ones wait for a person. In Cotool, listing Okta sign-ins runs automatically, while isolating a host in Defender or suspending an Okta user needs an approval in the app or Slack, and one click turns an agent read-only. Pair approvals with soft deletes or staged changes, like the ones Railway now applies to its API.
  • Restrict egress. An agent that can only reach the services it needs can't send data to an arbitrary address, whatever its weights were trained to do. Cotool runs each agent in an isolated Linux sandbox with outbound traffic restricted per agent, which is the control that matters most against the planted-behavior risk.
  • Version and review agent behavior. Keep agent instructions and tool grants in Git, where a change gets a pull request and can be rolled back. Cotool stores agents and skills as YAML in Git this way, and an AI judge grades every run against acceptance criteria.
  • Log and detect on agent actions. Check Point's Aaron Rose, quoted by ITPro, put it this way: an agent in production needs "its own discrete account, its own least privileged entitlements, its own behavioural baseline, and its own real-time audit trail." Cotool's detection page shows that applied to other people's agents: a threat hunt on developer AI sessions produced two detections, one for a hijacked AI agent session and one for a coding agent pivoting to credentials after untrusted instructions, and every detection run is logged with its tool inputs and outputs.

Mapped onto the PocketOS incident, those controls each break the chain at a different point. A token scoped to the staging environment couldn't have reached the production volume, a volumeDelete call labeled state-changing would have waited for a person, and a soft delete would have left a 48-hour undo. None of them depends on the model declining the request, which is why they hold up whether the model is aligned, abliterated, or backdoored.

When defenders want fewer refusals

The legitimate reason to want a model that refuses less is defensive work that looks like offense. During its July 2026 incident, Hugging Face needed to analyze more than 17,000 attacker events, and hosted frontier models blocked the requests because their guardrails "cannot distinguish an incident responder from an attacker." It switched to GLM-5.2, a standard open-weight model, on its own infrastructure, which also kept attacker data and credentials inside its environment.

That points to a better option than abliteration. A capable open-weight model from its original lab, run on infrastructure you control or through a provider you've vetted, sidesteps hosted guardrails without running weights someone else edited. Cotool Router takes that approach: when a provider refuses a defensive task or goes down, it moves the work to another provider or an open-weight model such as GLM 5.2, DeepSeek V4 Pro, or Kimi K3. Which AI models are best for security work compares how they score on defensive tasks, and Claude vs ChatGPT for cybersecurity covers each provider's access rules for cyber work.

  • Refusal direction: the single direction in a model's residual stream that Arditi et al. found controls whether it refuses.
  • Weight orthogonalization: editing a model's weights so no layer can write along a given direction, which makes abliteration permanent.
  • Jailbreak: getting a model to ignore its safety training through prompts, rather than by changing its weights.
  • Sleeper agent: a model trained to behave normally until a trigger appears.
  • Agentic offensive framework: attack tooling that puts a model in charge of security tools. See what ARTEX is and what an AI SOC is for the defensive side.

FAQ

What does abliterated mean?
Abliterated describes a language model whose refusal behavior has been removed by ablating, or zeroing out, the internal direction that makes it refuse. The word blends "ablate" and "obliterate." Abliterated models usually appear on Hugging Face with "abliterated" in the name, such as FailSpy's Llama-3-70B-Instruct-abliterated.
Is it obliterated or abliterated?
The term is abliterated, not obliterated, although "obliterated models" is a common misspelling in searches. An abliterated model is an open-weight model edited to stop refusing requests. Both spellings refer to the same technique, which Maxime Labonne's Hugging Face article popularized as "abliteration."
Are abliterated models safe to use?
Abliterated models are not unsafe to download by nature, but they remove one layer of protection and add supply chain risk. The weights come from whoever edited them, often an individual uploader, and the editing typically degrades quality. Safety then rests entirely on the controls around the model: scoped credentials, approval for destructive actions, egress limits, and logging.
Can an abliterated model contain a backdoor?
An abliterated model can contain a backdoor if whoever modified the weights also trained one in, and research shows such backdoors can be hard to remove. Anthropic's Sleeper Agents paper trained models that insert exploitable code when a trigger appears and found the behavior persisted through standard safety training. Abliteration itself does not add a backdoor, but you are trusting the editor as much as the original lab.
Why would a security team want a model that refuses less?
Security teams want a model that refuses less because investigating an attack means sending attack commands, exploit payloads, and malware to the model, which hosted models' guardrails can block. Hugging Face hit this during its July 2026 incident and switched to an open-weight model on its own infrastructure. A standard open-weight model or routing across providers addresses that without running weights someone has edited.

Sources

  1. Refusal in Language Models Is Mediated by a Single Direction (Arditi et al., arXiv) · accessed November 16, 2026
  2. Refusal in LLMs is mediated by a single direction - LessWrong · accessed November 16, 2026
  3. Uncensor any LLM with abliteration - Maxime Labonne, Hugging Face · accessed November 16, 2026
  4. failspy/llama-3-70B-Instruct-abliterated model card - Hugging Face · accessed November 16, 2026
  5. Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (Qi et al., arXiv) · accessed November 16, 2026
  6. LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B (arXiv) · accessed November 16, 2026
  7. Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (arXiv) · accessed November 16, 2026
  8. Deceptive LLMs that persist through safety training - Anthropic · accessed November 16, 2026
  9. Pickle Scanning - Hugging Face Hub docs · accessed November 16, 2026
  10. Data Scientists Targeted by Malicious Hugging Face ML Models with Silent Backdoor - JFrog · accessed November 16, 2026
  11. Your AI wants to nuke your database. Guardrails fix that. - Railway · accessed November 16, 2026
  12. Nine seconds was all it took for an AI agent to wipe a startup's database - ITPro · accessed November 16, 2026
  13. Security incident disclosure, July 2026 - Hugging Face · accessed November 16, 2026
  14. AI SOC Without the Black Box - Cotool · accessed November 16, 2026
  15. Introducing Cotool Router - Cotool · accessed November 16, 2026
  16. AI Detection Engineering with Python Detections - Cotool · accessed November 16, 2026

Keep reading

Attackers are scaling with tokens

Cotool helps defenders operate at machine speed. See how security teams are scaling Detection & Response.

Book a demo