Sovereign cybersecurity, within everyone's reachSovereign security for allSOC AI on your servers · free edition Request a demoDemo
Sovereign SOC AI · on your servers

Every action arbitrated. Every decision traced. Sovereign throughout.

LUAXERR keeps your infrastructure under control and explains every verdict: why an alert is kept, set aside, or needs checking. Every decision is written into a secured journal that lets you verify the records have not been altered. With the Enterprise edition, you also get a signed audit dossier, verifiable independently of LUAXERR. Your data stays in your infrastructure. Your decisions stay verifiable.

PLCs, cameras, building management, refrigeration plant, medical devices: LUIOTIX OT/IoT discovers and monitors what is plugged into your industrial network, listening only. Not one packet sent to your equipment, no data sent out, and it can be checked at your firewall.

  • 0 outbound calls
  • Free with no time limit
  • Ed25519 + ML-DSA-65 signed dossier, Enterprise
  • Published error rate
LUAXERR, demonstration in five scenes, without sound. An alert comes in and stays in your infrastructure. The local model returns a verdict and explains it. Your analyst approves the escalation. With the Enterprise edition, the audit dossier is signed twice, with Ed25519 and ML-DSA-65. If the model is unavailable, the alert is kept and escalated.

Demonstration replayed in a loop. The real processing runs on your infrastructure.

Frameworks addressedNIS2DORAISO/IEC 27001MITRE ATT&CKIEC 62443NIST CSFCRAISO/IEC 27090ANSSI

The five pillars

Guarantee and verification

SOC continuity and resilience

Detecting silent security sources, exercising degraded capabilities, and keeping a local fallback detection for when some paths stop answering.

Fallback detection →

Auditability and technical evidence

Chained log, signed audit file, traceability of decisions, and an explicit statement of what was not established.

The evidence →

Defence in depth around the AI

Deterministic controls against invented facts, contradictions and injections, with human validation and no automatic closing of alerts.

The guardrails →

Local control of operations

Designed to run without depending on outbound connections, data kept inside your environment, and an offline verification procedure.

Where the AI reasons →

Visibility into blind spots

Alerts cross-referenced with asset coverage, vulnerabilities, hardening findings and third-party relationships.

The blind spots →

How it works

Five steps. Humans decide, the machine proves.

The verdict informs the decision. The final analysis stays in your teams' hands.

Your SIEM

Connect

The alert becomes one pivot object, whatever its source: your SIEM, any device that emits CEF, or your EDR directly.

Automatic
  • Splunk
  • Microsoft Sentinel
  • QRadar
  • Elastic
  • Wazuh
  • Google SecOps
  • CEF
  • OCSF
  • EDR · IDS · IPS · NGFW direct, Enterprise edition

The alert arrives as is

Guardrails

Frame

Deterministic checks, outside the model: injection attempt, decoy touched, burst from a single actor. They can only escalate an alert, never silence it.

Automatic · deterministic

Local model

Judge

A named, versioned model running on your infrastructure returns a verdict with its reasoning, its confidence and the ATT&CK technique.

Automatic · local

Human approval

Approve

The tool proposes, you decide. No alert dropped on its own, no rule changed without you, no heavy action without approval.

Human

Signed proof

Prove

Every verdict carries the model and the prompt fingerprint behind it. In the Enterprise edition, the period dossier is signed twice, with Ed25519 and ML-DSA-65, and states what it does not establish.

Automatic · signed

0 data out · everything happens on your servers

Server room, racks lit in blue

Where the AI reasons

The model runs on your own instance.

On your infrastructure, or on a dedicated instance we operate for you, in France. Never on a shared platform: there is no common database and no shared console, so there is no path along which your data could meet anyone else's.

0bytes

Network egress, verifiable at your firewall

320cases

Control set the model was never trained on. Public measurement method

24h

NIS2 early warning tracked to its deadline, authority named

Hybrid

Signed audit dossier, Ed25519 and post-quantum (ML-DSA-65), verifiable offline without us

SHA-256

Prompt fingerprint written into every verdict, which replays

Transparency

Measured results, presented with their limits

We do not ask you to believe in our performance. We give you the means to check it.

Measured on 320 alerts the model was never trained on

Hand-written alerts, 160 of noise and 160 of attacks. Enterprise edition, model fine-tuned on 30 September 2026, measured on 10 November 2026.

Noise alerts recognised144 of 160
Attacks escalated by the model alone137 of 160
Median time per alert, on a graphics card2.3 s

These results describe the performance observed under the conditions of the protocol. They are not a guarantee that no error occurs in real conditions. No alert is closed automatically, the decision stays with your analyst.

See our measurement method

Two verdicts, as the model returns them

FALSE POSITIVE · 0.88

“RDP connection by the administration account adm.reseau from the official bastion to the domain controller SRV-AD-01, within the maintenance window opened by ticket CHG-4521: planned operation.”

action: no_action · prompt: 2170509b86ae · model: fine-tuned Mistral-7B v0.3 · translated from French

TRUE POSITIVE · 0.90

“Event 1102 on SRV-APP-07 at 02:58, followed by the creation of a local account and its addition to the Administrators group, with no ticket and outside the change window: classic post-exploitation technique after a takeover.”

action: escalate · prompt: 2170509b86ae · model: fine-tuned Mistral-7B v0.3 · translated from French

Each verdict comes with its reasoning, the prompt fingerprint and the model name.

Check

Run the measurement again

Our measurement method is public. A set of 60 alerts is sent on request, so that you can run the measurement again on your machine with the free edition, then on your own alerts.

See the method →
Safety of the system

A model error does not mean a lost alert

The model neither closes nor deletes any alert. Controls written in code, independent of it, automatically escalate a booby-trapped alert, a touched decoy or a burst.

The guardrail →
Before you choose

Four questions to ask any vendor

  • On how many cases is the model measured, and had it seen them before?
  • Is the measurement method published, and can I run it again?
  • How many attacks did the model misjudge?
  • What happens to a misjudged alert?

Our answers are on this page and in our measurement method.

Going further

The detail, where you need it

Six questions come up in every conversation. Each has its own page, with its diagrams, its figures and its limits.

Where the AI reasons

Your alerts handled at the source

The model runs on your infrastructure. No log is exported, and outbound traffic can be checked at your firewall.

An alert's journey →
The blind spot

The alerts you never receive

An EDR gone quiet, an asset outside the inventory, a perimeter never audited: LUAXERR names what has stopped talking, with no extra agent and no extra collection.

Source continuity →
Auditability

Your auditor's questions

Which model, where it reasons, whether a decision can be replayed, what the error rate is: each answer can be checked, and the signed dossier also states what it does not establish.

The four answers →
Industry, OT & connected devices

The plant floor too, offline

LUIOTIX listens passively on a mirror port. No packet sent to your controllers, no data leaves the site, and a survey signed twice.

Discover the OT/IoT probe →
What is different

The difference can be verified

Where the AI reasons, what has stopped emitting, the error rate, the proof for the auditor: six hard questions, and what LUAXERR does about each.

The full table →
Deployment

On your servers, or as sovereign SaaS

Same code, same signed dossier, same published error rate. What changes is who holds the machine.

Compare the two →

Offers

Which solution fits your situation?

Four situations, four answers. And three things that do not change in any of the four: the deterministic guardrail can only raise an alert, never set one aside; no alert is closed without an analyst; the audit dossier is verified offline, without us.

You are starting out, or working alone

Free edition · on your servers · with no time limit

Local triage and the anti-injection guardrail, the same ones as in the Enterprise edition. Nothing is crippled to make you pay.

What is free →

You have a security team

Enterprise edition · licensed · one to ten analysts

The full pipeline: triage, hunting, compliance, proposed remediation. Signed auditability dossier, and NIS2 or DORA declarations tracked to their deadline.

Discover the Enterprise edition →

You have a plant, OT or connected devices

LUIOTIX, OT/IoT probe · passive listening · separate solution

A probe on the mirror port. Not one packet sent to your controllers, because an active scan on a production line is an incident, not an audit.

Discover the OT/IoT probe →

You run security for other companies

Managed services · one deployment per client

Each client inside its own perimeter. No data moves from one client to another, because no path exists for it to move through.

Write to us →

Frequently asked

What we get asked first

Do my logs leave my infrastructure?

No. The solution runs on your infrastructure, without depending on a remote AI service. Intelligence comes in as a signed pack, loaded offline. Check it at your firewall: that is a verification, not a statement.

Do I need a GPU?

Not to start, and here are the two figures to decide for yourself. On a standard processor we measure 170s per alert on average, 573s in the worst case. On a graphics card rented by the hour, the same chain returns its verdict in 2.3s, 74 times faster, and the worst case falls to 3.5s.

The card speeds things up, but the verdict does not replay identically on it. The hardware choice therefore also depends on your audit requirements. Sizing is worked out on yours.

I do not have a SIEM yet.

Wazuh is a free option to centralise your events, and it feeds LUAXERR directly. If you already run Splunk, Microsoft Sentinel, QRadar, Elastic or Google SecOps, integration depends on the product and its version; it is confirmed when we study your environment.

What is your false positive rate?

We measure it on 320 alerts the model was never trained on, 160 of noise and 160 of attacks. In the measurement of 10 November 2026, 144 noise alerts out of 160 were recognised and 137 attacks out of 160 were escalated by the model alone.

These results describe the performance observed under the conditions of the protocol. They are not a guarantee that no error occurs in real conditions: that is why no alert is closed automatically.

The measurement conditions and their limits are described in our measurement method.

All ten questions, with their answers →

Scope

What LUAXERR does not do

  • It is not a detection engine. It replaces neither your SIEM nor your rules. It takes over what they produce.
  • It is not an oracle. The model gets things wrong, and the evaluation set shows it with numbers.
  • It is not real-time. Scoring takes from a few seconds to a few minutes depending on your hardware.

What it does, it does verifiably.

Contact

See LUAXERR judge an alert from your environment

A demo on your context, not an idealised scenario. We answer in person, within two business days.

This site uses cookies. Technical cookies are required for the site to work and stay on at all times. Functional and audience-measurement cookies wait for your consent.

Customise