# Open models for security research, in the EU

URL: https://munito.ai/solutions/security-research

A guarded model refuses the analysis of malware, phishing and exploit text. Name the open model your security team needs, and we run it on capacity of your own.

## When the model refuses, the analysis stops

A malware sample. A phishing corpus. An exploit write-up from a disclosure you are triaging. A guarded model reads the words and declines, and your analyst goes back to doing it by hand.

That refusal is correct for a consumer chat product and wrong for a security team. Munito AI runs open models on capacity you reserve, so you can name the model the work needs.

## Two models we run on request

These are not in the public catalogue and they carry no posted price. We provision each one as a [dedicated endpoint](https://docs.munito.ai/dedicated-endpoints) for your organization alone, in the region you pick, once the work is agreed.

| Model                                   | Base                     | How you get it |
| --------------------------------------- | ------------------------ | -------------- |
| GLM-5.3-Flash, abliterated               | `zai-org/glm-5.3-flash`  | On request     |
| DeepSeek-V4.1-Flash-1017, abliterated    | `deepseek-ai/deepseek-v4.1-flash-1017` | On request |

Both are community builds of open weights with the refusal direction removed. We serve the base models already, so the capacity, the engine and the API are the ones we run every day.

**Write to [support@munito.ai](mailto:support@munito.ai)**, or open Support in the [console](https://console.munito.ai), with the organization, the work and the volume. We answer with the capacity, the price and the contract.

## Or name a model of your own

A dedicated endpoint runs any open model, not only these two.

| What you bring        | What we do                                                  |
| --------------------- | ------------------------------------------------------------ |
| The open model you need | Place it, provision it, and serve it on reserved GPUs       |
| Your request rate and latency target | Size the capacity for it                     |
| Your region           | Run it in the EU, the US, your own cloud, or on-premises     |
| Your use case         | Agree it in the contract before anything is provisioned      |

You call it with the same API and the same key as the shared models, through the OpenAI or the Anthropic interface.

## The workloads this unblocks

- **Malware and script triage.** Summarise what a sample does, in plain language, for an analyst who then decides.
- **Phishing corpora.** Classify, cluster and describe campaigns from real messages, not sanitised ones.
- **Disclosure triage.** Read an advisory or a proof of concept and extract the affected versions, the conditions and the mitigation.
- **Detection engineering.** Draft and explain rules against real artefacts.
- **Red-team support.** Reason over your own engagement scope, for engagements you are authorised to run.

The analyst stays in the loop on every one of them. The model reads and drafts; a person decides and signs.

## Detonate it on a machine that is not yours

Reading a sample is one risk. Running it is another. A [sandbox](https://munito.ai/sandboxes) is an isolated cloud computer started in about a second and thrown away afterwards, and [its network policy](https://docs.munito.ai/sandboxes/networks) is yours to set, which for a sample means no egress at all.

One sandbox per artefact, no shared state, billed per second from $0.0073 an hour.

## Samples are the data you least want kept

Customer artefacts, victim data, and sometimes the contents of a breach. Zero data retention is the default here: we keep no prompts and no outputs, we train on nothing you send, and we run no human review and no content moderation over what you send.

Pin the region with the `api.eu` host and read `munito-region` off the response. The DPA applies with no signature, at [legal.munito.ai/dpa](https://legal.munito.ai/dpa), and the [sub-processor list](https://trust.munito.ai/subprocessors) names every company by region.

For work that may not leave your perimeter, the platform runs inside your own cloud account or on-premises. See [Enterprise](https://munito.ai/enterprise).

## How we decide

A dedicated endpoint for this work is agreed in a contract, not clicked. Tell us the organization, the work and the model, and we scope it with you. We serve authorised defensive and research work: threat analysis, detection, triage and red-teaming you are engaged to perform. We do not serve a request to build or operate an attack.

A build with its refusals removed drops them everywhere, not only on the analysis you want. So the gate is the customer and the contract, not the weights. Your own policy, your own logging and your own review sit on top, and we expect them to.

## Start

- [Talk to an engineer](https://munito.ai/contact) about the model, the volume and the region
- [Read how a dedicated endpoint works](https://docs.munito.ai/dedicated-endpoints)
- [Isolate a sandbox from the network](https://docs.munito.ai/sandboxes/networks)
- [Bring the paperwork to your review](https://munito.ai/security-review)