CyberThor by SquareX ApS  ·  your data stays with you

> Your own security AI.
Private by design.

CyberThor is a family of domain-specific LLMs fine-tuned for one job: reading what your systems report and telling you what it means and what to do about it. It works inside your own environment, never sends a log line to an outside AI service, and we set it up and look after it for you.

No outside AI service Set up and managed for you You approve every change
  CyberThor-Sentinel-8B · case AG-7f3c21 ● local
event-01 sshd: Failed password for root from 203.0.113.47
event-02 sshd: Invalid user admin from 203.0.113.47
event-03 summary: 71 failed logins, 5 usernames, 0 accepted
classification brute_force_attempt
severity medium
evidence event-01, event-03
assessment Password guessing from one source, no successful login.
proposed_action block_source_ip 203.0.113.47
status waiting for your approval
0bytesof your security data sent to an outside AI service
2modelsCyberThor-Sentinel-8B and CyberThor-Sentinel-4B
1fileis the whole model. We install it and keep it up to date
5actionson a fixed menu. It proposes; it never runs free

Why private

Security data is the last thing
you should send to someone else's AI.

Logs, alerts and scan results describe exactly where you are weak. CyberThor analyses them where they already are: inside your network.

Private by design

CyberThor works inside your environment and calls no outside AI service. What happens in your network stays in your network.

Fully under your control

You own the installation. You decide which actions are automatic, which need approval and which are never allowed. Every decision is written to an audit log you keep.

Easy to adopt

No new data centre and no special hardware. We size it, install it, keep it updated and look after it, so your team does not have to become AI operators.

How it works

From raw finding to a decision you can sign.

CyberThor does not replace your scanners or sensors. It reads what they found, the way an experienced analyst would, and hands you a clear, structured answer.

Collect

The agent reads what your hosts report: pending updates, failed services, login attempts, scanner output. Read-only.

Look up

It adds facts from a local threat-data copy: known-exploited vulnerabilities, exploit likelihood, vendor notices.

Judge

CyberThor classifies the finding, rates severity, cites the evidence it used and explains its reasoning in plain language.

Propose

It suggests one action from a fixed menu. Your policy decides: automatic, needs approval, or never.

Vulnerabilities and patching

Which finding to fix first, and why. Security update or routine update. Reboot needed or not.

Logins and host health

Brute-force attempts, suspicious successful logins, failed services, full disks, pending reboots.

Windows, Linux, macOS and cloud

Recognises known attack techniques in process activity and in AWS, Azure and GCP events, and tells them apart from everyday administration.

Model family

One model family. Two sizes.

Both are domain-specific LLMs fine-tuned for security triage. We help you pick the right one and set it up for you.

model registry
$ cyberthor models --list
MODEL                    PARAMS   QUANT    SIZE      RUNTIME
CyberThor-Sentinel-8B    8.03B    Q4_K_M   4.9 GB    CPU / GPU
CyberThor-Sentinel-4B    4.02B    Q4_K_M   ~2.5 GB   CPU / GPU
Full model

CyberThor-Sentinel-8B

The best judgement and the deepest security knowledge. For your main environment.

  • Parameters8.03 billion
  • QuantizationQ4_K_M (4-bit GGUF)
  • Model file4.9 GB
  • Memory at inference7.9 GB
  • Latency per finding, CPU onlyabout 12 s
  • Recommended host8-core CPU, 16 GB RAM
Light model

CyberThor-Sentinel-4B

Half the parameters, about twice the speed. For smaller sites and tighter budgets.

  • Parameters4.02 billion
  • QuantizationQ4_K_M (4-bit GGUF)
  • Model fileabout 2.5 GB (target)
  • Memory at inferenceabout 4 GB (target)
  • Latency per finding, CPU onlyabout 6 s (target)
  • Recommended host4-core CPU, 8 GB RAM (target)

Sentinel-8B figures were measured in our lab on an AMD Ryzen 7 9700X using 8 threads. Sentinel-4B figures marked "target" are design targets and will be replaced by measurements. Both are faster with a GPU.

Under the hood

Built like a model, shipped like a product.

For the engineers in the room: what CyberThor is made of, and how it is served.

architectureDecoder-only transformer LLMs, built on open foundation models
fine-tuningQLoRA (r=16, α=32) on 12,919 labelled security cases; loss on the answer tokens only
coverageVulnerabilities, patching, authentication, host health, Windows / Linux / macOS and cloud activity
inferenceQuantized GGUF weights (Q4_K_M) on llama.cpp; CPU or GPU; no external calls
interfaceOpenAI-compatible chat-completions API on a local endpoint
decodingTemperature 0; one JSON object per finding, validated against the supplied evidence IDs
action spaceFive fixed actions; every proposal re-checked against observed facts, then policy tiers
evaluationHeld-out cases split by incident, a sealed final test set, and a public cyber-knowledge benchmark

Control

It advises. You decide.

An AI that can change your systems has to be boring in exactly the right places. CyberThor is built so that even a mistaken or manipulated answer cannot do damage.

AutomaticOnly the actions you have switched on, for example installing a security update.
Needs approvalWaits in a queue until a person approves, for example blocking an address.
NeverAnything outside the fixed menu. The model cannot write commands.

A fixed menu of actions

Install a security update, schedule an update that needs a reboot, restart a service, block an address, or raise to a person. Nothing else exists.

Every proposal is double-checked

Before anything runs, the agent verifies the proposal against what it observed itself. Text hidden in a log cannot trigger an unrelated action.

Report-only by default

Every installation starts by only telling you what it would do. A full audit log and a one-step kill switch are always there.

Straight talk

What CyberThor is, and what it is not.

What it is

  • Our own security AI, built on open foundation models and trained by us for security triage.
  • A tireless first reader of findings that explains itself and cites its evidence.
  • Software you control, updated when you choose.

What it is not

  • Not a detection sensor. It interprets what your tools capture.
  • Not a replacement for your security team or a guarantee against attack.
  • Not autonomous. People stay in charge of anything that matters.

Services around it

People who test, validate and stand behind the result.

CyberThor keeps watch. The specialists at SquareX ApS find the weaknesses, prove the fixes and prepare you for the bad day.

Security Baseline

A full picture of your exposures across servers, cloud, identities and applications, with a 30/60/90-day plan.

Security Testing

Penetration tests of networks, web applications and APIs, cloud reviews and identity assessments. Always in writing, always in scope.

Continuous Security

CyberThor set up and operated in your environment, with scheduled scanning and a monthly review by a specialist.

Recovery & Resilience

Ransomware readiness, backup and restore checks, recovery plans and exercises that are actually run.

See it on your own data.

Start with a report-only pilot. We set CyberThor up in your environment, it changes nothing, and it shows you what it would have done. You judge it for yourself.

Contact us