Open-weight models, rebuilt for security work

We select leading open-weight models adapted at the weight level for cybersecurity tasks by specialised developers. They engage with real security work, and every model runs on its own dedicated GPU as a private endpoint.

Adapted at the weight level

A system prompt or a filter sits on top of a model. We change the model itself, so the behaviour is part of the weights.

  1. 01

    Select

    We start from leading open-weight models, from compact to frontier scale.

  2. 02

    Adapt

    The weights are adapted and modified for cybersecurity tasks. The model itself learns to work with exploit reasoning, malware and attack technique.

  3. 03

    Deliver

    Every model ships as a Blue Team variant, and as a Red Team variant where its licence allows it, and runs as a private endpoint on your own dedicated GPU.

Four model classes, one for every job

Start small and move up when the work needs it. Every class is available as a Blue Team variant, and as a Red Team variant where the licence of the model allows it.

Class 01

Entry

Fast and affordable, for everyday assistance

Compact models that answer quickly. A good first instance for interactive work and high-volume, well-defined tasks.

Best for

  • Alert and log triage
  • Explaining CVEs, advisories and findings
  • Reviewing scripts, configs and small code snippets
  • Drafting report sections and ticket summaries
  • Interactive help during an assessment

Example models

  • Muse Glimmer 30B
  • Gemma 4 31B
  • Qwen3.8 27B

Access

Blue Team: No verification required

Red Team: Verified email and intended use

Class 02

Basic

The workhorse for real analysis

Large open-weight models with the depth for longer inputs and multi-part reasoning. Most teams start here.

Best for

  • Source code and repository review
  • Vulnerability analysis of a single component
  • Malware triage and reverse engineering support
  • Long reports, playbooks and runbooks
  • Reading screenshots and diagrams with vision-language models

Example models

  • GPT-OSS 120B
  • Nemotron 3.5 Super VL 120B-A12B

Access

Blue Team: No verification required

Red Team: Verified email and intended use

Class 03

Pro

Built for agents and demanding workloads

Stronger tool calling and longer chains of reasoning, made for agent harnesses and coding agents that run on their own.

Best for

  • Agentic workflows with many tool calls
  • Plugging into coding agents and security platforms
  • Audits across large codebases
  • Exploit chain reasoning
  • Automated recon and assessment pipelines

Example models

  • GLM 5.3 Flash
  • Qwen3.8 Flash Next
  • MiMo V2.6 Flash RL

Access

Blue Team: No verification required

Red Team: Full verification required

Class 04

Frontier

Frontier-scale research capability

The largest models on reserved capacity, for the hardest problems and for teams that need the highest ceiling.

Best for

  • Deep vulnerability research
  • Complex exploit and malware analysis
  • Long-horizon autonomous security agents
  • Cross-system threat research
  • Reserved capacity by individual agreement

Example models

  • GLM 5.3
  • DeepSeek V4.1 Flash 1017
  • Kimi K3

Access

Blue Team: No verification required

Red Team: Full verification, business customers only

Blue Team and Red Team variants

Each model class comes in two variants, tuned for defensive and for offensive security work.

Blue Team

Blue Team variant

Built for defensive work that has to engage with attack technique. These models understand active malicious code in depth, including its behaviour, exploit mechanics and intrusion tradecraft, without creating it. Where general-purpose APIs typically refuse or dilute the answer, they respond within defined limits.

  • Incident response
  • SOC analysis
  • Code review
  • Repository analysis
Red Team

Red Team variant

Built for authorised offensive security research. These models work with offensive technique directly, from vulnerability analysis and exploit reasoning to reverse engineering and malware analysis. They take on the work that general-purpose APIs decline. Red Team access needs a short description of your intended use, Pro and Frontier need full verification, and every use remains subject to the terms and applicable law.

  • Penetration testing
  • Vulnerability research
  • Exploit analysis
  • Reverse engineering
  • Malware analysis

Use is limited to authorised security work. See the acceptable use policy.

Request early access

Tell us about your team and the models you want to run. We will contact you as soon as your tier is available.