Cybersecurity LLMs

Abliterated models

Run abliterated LLMs for red teaming, penetration testing and cybersecurity work, all through one API on Opper.

3 abliterated models2 providers3 routes

Abliterated models on Opper

Each model needs approved access. Once your account is verified you call them with your usual Opper API key, and new abliterated models appear here as they go live. Every route runs with zero data retention, some with abuse monitoring.

by Morph

DeepSeek V4.1 Flash Abliterated is Morph's refusal-removed version of DeepSeek V4.1 Flash: a 1M-context multimodal reasoning model for security, red-team and agent-testing work.

Available on

Pricing from$0.30/ 1M input$1.20/ 1M output$0.03/ 1M cache read1Mcontext

Privacy

ZDR
Training
No
Region
US
by abliteration.ai

Abliterated Large v2 by abliteration.ai is GLM-5.3 with refusals removed: a 1M-context reasoning model for cyber, red-team and agent-testing work.

Available on

Pricing from$3.00/ 1M input$5.00/ 1M output$0.30/ 1M cache read1Mcontext

Privacy

ZDR
Training
No
Region
US
by abliteration.ai

Abliterated Model by abliteration.ai is a general-purpose refusal-removed model with image input and a 256K context, built for security, red-team and trust and safety work.

Available on

Pricing from$1.00/ 1M input$3.00/ 1M output$0.10/ 1M cache read262Kcontext

Privacy

ZDR
Training
No
Region
US

The open models behind them

Each abliterated model starts from an open-weight model. The edit removes its refusals and keeps the rest.

Abliterated modelMade byBased on
DeepSeek V4.1 Flash AbliteratedMorphDeepSeek V4.1 Flash by DeepSeek
Abliterated Large v2abliteration.aiGLM-5.3 by Z.ai
Abliterated Modelabliteration.aiQwen3.5 family by Qwenabliteration.ai does not name the exact checkpoint

What an abliterated model is

Abliteration finds the direction in a model's activations that makes it refuse and removes it from the weights. Nothing else is retrained, so the model keeps the base model's knowledge and skills and stops declining requests.

An uncensored model is usually fine-tuned on new data to loosen its rules, which can shift its quality. Abliteration edits the refusal behaviour directly, so results stay close to the base model.

Abliterated vs uncensored vs base models

Three versions of the same open-weight model, and how they differ.

ComparedAbliteratedUncensored fine-tuneBase model
How it is madeThe base model with its refusal direction removed from the weightsThe base model fine-tuned on new data to loosen its rulesThe lab's own release, with its safety tuning
RefusalsFollows through on requests the base model declinesFewer, depending on the fine-tune dataDeclines security and other sensitive requests
QualityClose to the base model, since nothing else is retrainedCan shift with the fine-tuneThe reference

Call abliterated models through one API

Once your access is approved, every abliterated model here works through Opper's OpenAI-compatible API with your usual key, so moving between them is a change of model id.

Let your coding agent set it up

Paste this into Claude Code, Codex or Cursor.

Or call it yourself

import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.OPPER_API_KEY,
baseURL: "https://api.opper.ai/v3/compat",
});
const completion = await client.chat.completions.create({
model: "morph/morph-dsv41flash-abliterated",
messages: [{ role: "user", content: "Hello" }],
});
console.log(completion.choices[0].message.content);

How access works

Abliterated models stay off every account until we have reviewed who wants them and why.

  1. 1. Get your API key

    Sign up for Opper without a credit card. Access is enabled on that account.

  2. 2. Request access

    Email us who you are, your organization and what you plan to test. No mail app? Write to hello@opper.ai

  3. 3. We review it

    Every request gets a manual identity and use-case check. Once approved, the models work with your API key and SDKs.

Responsible use · Abliterated models are for defending systems, so access comes with a few plain rules.

  • Your own systems only. Use them only on systems you own or have written permission to test.

  • A verified account. Access is tied to a verified person or organization, never an anonymous key.

  • Misuse ends access. Breaking these rules ends access to abliterated models on your account.

Who you contract with

Company
Opper Technology AB, Sweden, reg. no. 559446-3720
Certification
Gateway
AWS Stockholm, no prompts stored by defaultSecurity overview
Pricing
Processor
Opper, for every modelSub-processor list
DPA
One DPA covers the whole catalogueRead the DPA

Frequently asked questions

What is an abliterated model?

+
An abliterated model is an open-weight LLM with the direction that makes it refuse removed from its weights. Nothing else is retrained, so it keeps the base model's knowledge and skills.

Why does Opper review access to abliterated models?

+
Because they do not refuse, we want to know who uses them and for what. Every account is reviewed by hand before the models are enabled, so they are used to test and defend systems.

How do I get access?

+
Sign up for an Opper account, then email hello@opper.ai with your organization and what you plan to test; the Request access button on this page opens that email for you. We check your identity and use case, case by case, and enable the models on your account. You then call them with your usual API key.

What do Opper and the model hosts keep from my requests?

+
Those are two separate answers. Opper stores only usage metadata such as model, tokens, latency and cost by default, tracing is a rule you turn on with retention from 1 to 30 days that projects can shorten to zero, and Opper never trains on your data. On the hosts' side, every route runs with zero data retention, some with abuse monitoring. Each model's page shows the full data handling per route, including caching.

Where do abliterated models run on Opper?

+
They run in the US, hosted by Morph and abliteration.ai. Input runs from $0.30 to $3.00 per million tokens.

How is an abliterated model different from an uncensored model?

+
An uncensored model is usually fine-tuned on new data to loosen its rules, which can shift its quality. Abliteration edits the refusal behaviour directly, so results stay close to the base model.

Do abliterated models perform worse than the base model?

+
Usually only slightly. Abliteration removes one direction from the weights and retrains nothing else, so results stay close to the base model. Each model's page shows its benchmarks where they are published.

What is a cybersecurity LLM?

+
A language model used for security work: testing, analysing findings, writing detections and reports. It can be a general model, a security-tuned one, or an abliterated one, which does not refuse security requests. The models on this page are abliterated.

Can I use an LLM for penetration testing?

+
Yes, on systems you own or have written permission to test. Mainstream models often refuse parts of a test; abliterated models follow through, which is why Opper reviews access to them per account.

Read this list from an agent

This list as data. The same models, base models, prices and regions as markdown at opper.ai/models/abliterated.md.

The whole catalogue. The model catalogue is public JSON at https://api.opper.ai/v3/models, no API key needed. Filter it with query parameters, for example ?type=llm&limit=3, and page through it with limit and offset. Use it when you can fetch a URL but cannot connect to an MCP server. Coding agents that use MCP can connect to https://api.opper.ai/mcp. Two of the server's tools need no account and no sign-in: list_models searches the model catalogue by name, type, provider or capability, and get_guide returns short setup guides. Everything that touches your account needs you to approve the connection in the browser first. About the MCP server

Test your defences before attackers do

Get your API key, request access, and once we have approved it call abliterated models through the same Opper API as every other model.