Praetorix AE · Artificial Expert

An expert system that runs on your own infrastructure

Vulnerability verification with proof of exploitability, decompilation of binaries with no source, and control of your communication perimeter — without binaries, findings or mail ever leaving your network.

The model runs at your site Verdicts backed by evidence Delivered appliance Perpetual licence
Principle

We only claim what we can prove

In security analysis a language model hallucinates in a way that is worse than silence: an invented vulnerability burns a day of the team's work, a missed one kills the audit. That is why in AE the model never speaks to the user directly.

Layer 1

Heuristics

Disassembler, static analysis, signatures, symbolic execution, reachability checks. Deterministic, fast and explainable — but brittle, with weak recall on patterns it has never seen.

Role: produces candidates and a hard factual baseline.
Layer 2

Model

Generalises, names things, reconstructs the intent of code and the connections heuristics cannot see. Its statements, however, are unverifiable on their own.

Role: produces hypotheses — never conclusions.
Layer 3

Oracle

Recompilation and semantic comparison, proof executed in an isolated sandbox, a verifiable fact from DNS and your records. Decides deterministically whether the hypothesis holds.

Role: decides whether a claim leaves the system.

Every finding gets one of three states

A tool with no way to say “I don't know” is lying. The third state is therefore a deliberate part of the design, not a gap.

CONFIRMED

The oracle passed and evidence exists — the recompiled code behaves identically, the proof actually ran in the sandbox, the fact matches your records. The output stands up as audit evidence.

REFUTED

The oracle showed the finding does not hold — the vulnerable code is unreachable, the version does not match, the configuration rules the problem out. The false positive is closed with a reason.

UNPROVEN

The oracle reached no conclusion. The finding goes to a human together with what was established and where the analysis hit a wall. No silent conclusions, no filling gaps with guesses.

What AE does

Seven jobs a senior engineer does by hand today

Every engine rests on the same triad — heuristics, hypothesis, verification. They differ in their input and in what serves as the oracle.

AE-VER — vulnerability verification

A finding from a scan or pentest is checked for version, configuration and reachability; the model proposes an exploitation path and a proof; that proof runs in an isolated sandbox against a clone of the target. The result is not a score but proof or refutation — and a recalculated residual risk.

AE-DEC — decompilation

A binary with no source — vendor firmware, a legacy component, a suspicious executable. The model reconstructs names, types and the intent of functions; the oracle recompiles the result and compares it with the original. The output marks what is verified and what is an estimate.

AE-COMM — communication perimeter

An inventory of channels, control of what comes in and what goes out. Payment-detail fraud, look-alike partner domains, attachments detonated in a sandbox, data leaving through an unapproved channel. In detail below →

AE-PATCH — patch analysis

What a vendor patch actually fixes and whether it concerns you. A binary diff before and after, matched against your component inventory. The answer to the question that costs hours with every bulletin — must we patch now, or are we not affected.

AE-TRIAGE — alert triage

First-line triage of security alerts through the same harness. A claim about an alert must agree with an independent telemetry source; otherwise it is marked unproven and reaches the analyst with the context already gathered.

AE-CODE — code review

Review of your own code against your conventions and your libraries, not against a generic template. A finding counts only once there is a test case that reproduces the defect.

AE-DOC — regulatory output

Translation of a technical finding into the language of cybersecurity law and its implementing rules, including entries in the statement of applicability and the risk analysis. The engineer does not write the documentation, they review it.

AE-COMM

Control of the communication perimeter

Communication is the largest and worst-mapped attack surface a company has — and the only channel that crosses the organisation's boundary in both directions every day. AE-COMM handles it on three levels.

The perimeter

What we actually have
  • Channel inventory — mailboxes and domains, mobile channels, file-transfer services, shared storage, API integrations. For each: who owns it and what data travels through it.
  • Shadow channels — the ones in daily use that appear in no policy. They typically surface here and nowhere else.
  • Authentication posture of your own domains — including unused ones. An unused domain without a strict policy is a ready-made ramp for fraud in your company's name.
  • Counterparty posture — the same suppliers you exchange invoices with. Their weakness is your risk, not theirs.
  • Communication graph — who talks to whom, how often and with what class of data. That is what reveals which relationships are payment-critical.

What comes in

Inbound control
  • Payment-detail fraud — a changed account number against the partner's history. The finding stands only after a mismatch with your records, not because of the tone of the message.
  • Look-alike domain — similarity to partner domains, age of registration, registrar. The assessment is factual, not impressionistic.
  • Failed sender authentication — deterministic evaluation; the model is not even consulted.
  • Attachments — type, macros, signature; a binary attachment goes to sandbox detonation and to AE-DEC.
  • Links — the redirect chain is unrolled in the sandbox and compared with your own sign-in page.
  • Pressure patterns — urgency and process bypass aimed at a specific role.

What goes out

The direction that gets ignored
  • Class of outbound data — personal and health data, contractual and pricing information, source code, credentials. A separate layer classifies it, not the language model.
  • Secrets in attachments — keys, tokens, database exports.
  • Wrong recipient — sending to a domain resembling a partner's, or outside the circle that relationship has any reason to reach.
  • Volume and channel — a bulk export through an unapproved service, which otherwise disappears in the mail log.
  • Failed delivery — permanent failures are not noise but an indicator that part of your communication is quietly not arriving.
Harness

The classifiers around the model

The model is a replaceable part. What turns it into a usable tool is the envelope of classifiers in front of it and behind it — and a gate no change passes without evaluation.

PRE

Task scope

Decides whether a task belongs to AE at all and to which engine. Prevents the tool from degenerating into a general-purpose chat over company data.

PRE

Secret redaction

Keys, passwords and personal data are stripped before input reaches the model. The inputs are memory dumps, logs and configurations — material full of sensitive data.

PRE

Hidden-instruction detection

AE's inputs are attacker-controlled — strings in malware, lines in a log, message content. An attempt to embed an instruction for the model must be caught before the model sees it.

PRE

Task budget

Estimates cost and splits or refuses the task. Compute is a finite resource: it is planned, not spent.

POST

Anchoring in evidence

Every claim must point to a specific place — an offset in the binary, a line of code, a record. A claim without an anchor is discarded.

POST

Overclaim check

Catches the shift from “possible” to “confirmed” without the oracle ever running. This is the most common way AI tools lie.

POST

Confidence calibration

The model's score is converted into confidence calibrated against real historical accuracy. No “97 %” pulled out of the model.

POST

Cross-run agreement

The same task runs independently several times. Disagreement between runs means escalation to a human, not averaging the result.

GATE

Evaluation before deployment

No change to a classifier or the model reaches production without a run over an annotated set and a signature. A profile is either promoted, rolled back or rejected.

Praetorix HW

We supply the hardware

AE is not sold as software you install somewhere. It arrives as a sealed appliance, is plugged in and runs. The reason is not convenience but provability — if the client assembles the machine, nobody can attest what runs on it or whether the model weights ever left the node.

Praetorix HW-1

One entity, entry-level deployment
  • One compute unit in the 24–48 GB class
  • Engines: communication perimeter, triage, code review, regulatory output
  • Separate sandbox node included
  • Remote management and updates
I'm interested
Most common choice

Praetorix HW-2

Full operation for a single entity
  • Two units in the 80 GB class
  • Everything in HW-1 plus vulnerability verification and patch analysis
  • Sandbox for running proofs of exploitability
  • Integration with the vulnerability register and risk analysis
I'm interested

Praetorix HW-3

A group of several entities
  • Four units in the 80 GB class, long context
  • Everything in HW-2 plus decompilation of large binaries
  • Concurrent operation for several group companies
  • Priority support
I'm interested
BUILD

Two separate nodes

The compute node holding the model and the sandbox node are physically separate, not separated by configuration. Configuration separation can be undone by mistake — and an exploit running next to the model reaches the weights and your data.

BUILD

Sealed assembly

The component list and their fingerprints are part of the delivery note. Opening the chassis is detectable.

BOOT

Measured boot

Boot fingerprints are compared at every start. A mismatch is an event that gets acted on — not a quiet carry-on.

DATA

Encryption at rest

Disks and model weights alike. Keys are bound to the boot state, so a disk removed from the appliance is unreadable.

NETWORK

No outbound connectivity

By default no outbound traffic except the secured administrative channel. In the closed variant, not even that.

MGMT

Signed updates

Model, classifiers and firmware are signed and verified before loading. For closed environments there is a dedicated package on media — never “no updates at all”.

MGMT

Remote state verification

That an update really happened is verified by the fingerprint of the running version, not by an operator's report.

SERVICE

Disks stay with you

On a service replacement the appliance ships without its disks — they hold your data and the model. This belongs in the contract, not in a verbal agreement.

The appliance is sold or leased; the AE licence is bound to it by the same mechanism as Praetorix perpetual licences. Power draw and consumption are quoted up front — on the larger tiers it is a line item operations will notice.

Deployment

Three modes, depending on what you operate

CLOSED

No connectivity whatsoever

For operators of regulated services and industrial control environments. The appliance has no link to the outside world; updates arrive as a signed package on media and are verified before loading.

RECOMMENDED

At your site, managed remotely

The appliance runs in your network and we manage it over a secured administrative channel. Every unit has its own keys — nothing is shared between customers.

YOUR CLOUD

In your private environment

For organisations without their own hardware. It runs in your environment under your identity management; data stays within boundaries you define.

Scope of delivery

A core and separable modules

The core is mandatory; the remaining modules are bought according to what you actually need. The same principle as the Praetorix core — you pay for what you use.

AE1

Core

The classifier harness, profiles and their evaluation, evidence records, integration with monitoring. Mandatory in every delivery.

AE2

Vulnerability verification

The AE-VER engine and the sandbox for running proofs. Builds on the vulnerability register in Praetorix.

AE3

Decompilation

The AE-DEC engine with recompilation checking. For vendor firmware, legacy components and suspicious files.

AE4

Triage and patches

The AE-TRIAGE and AE-PATCH engines. Build on continuous monitoring and the component inventory.

AE5

Code review

The AE-CODE engine over the organisation's own code, with a reproducible test for every finding.

AE6

Regulatory output

The AE-DOC engine — turning findings into risk analysis, the statement of applicability and audit outputs.

AE7

Communication perimeter

The AE-COMM engine — channel inventory, control of inbound and outbound communication, including the impact assessment and template documentation.

FAQ

The questions we get most

Will our data leave our infrastructure?+

No. The model runs on the Praetorix HW appliance at your site. By default the appliance has no outbound connectivity except a secured administrative channel — and in the closed variant not even that. Binaries, source code, findings and mail are never sent anywhere.

How is AE different from a chatbot over company data?+

In AE the model never speaks to the user directly. It only produces hypotheses that a deterministic oracle must confirm — recompilation, a proof executed in the sandbox, or a verifiable fact from your records. A claim without evidence is not released. AE also has no open conversational mode and no right to act on production systems: it proposes an intervention, it does not perform one.

What does the three-state verdict mean?+

Every finding gets one of three states — CONFIRMED (the oracle passed, evidence exists), REFUTED (the oracle showed the finding does not hold) and UNPROVEN (the oracle reached no conclusion, the finding goes to a human). The third state is deliberate: a tool with no way to say “I don't know” is lying — and that is precisely why AI findings cannot be trusted.

Is communication control employee surveillance?+

No. The purpose is detection of cyber risk, not assessment of people — the tool has no performance-assessment function and will not be given one. The default mode processes metadata, not content; access to content requires a second approval and is recorded. Communication with legal counsel, trade-union and whistleblowing communication are excluded technically. A data protection impact assessment and template documentation for informing employees are part of the deployment.

Why do you supply the hardware?+

If the client assembles the machine, we cannot attest what runs on it or how it boots. With a delivered appliance the whole chain is provable — a sealed assembly with fingerprints on the delivery note, measured boot, and encryption bound to the boot state. For a security tool this, not raw performance, is what the audit examines.

How does AE relate to Praetorix and Vigil?+

Praetorix governs compliance — records, risks, the statement of applicability and audit outputs. Vigil provides continuous monitoring. AE is the third layer: the technical analysis that turns a finding into a verified conclusion. AE findings feed the vulnerability register and the risk analysis. It can also be deployed on its own.

Will AE replace our pentester or analyst?+

No, and it does not try to. It removes the routine — sorting false positives, first-round triage, transcribing findings into documentation — and hands over a substantiated basis. The decision stays with the human, and everything the oracle did not verify is handed over explicitly.

No-obligation request

Request a Praetorix AE demo

Fill in the form — we will open a tracked ticket and come back within 24 hours with a proposed date. We can build the demo around your own case: a finding from a scan, a binary with no source, or an overview of your communication channels.

  • No obligation, no cost
  • A tracked ticket with a reference number
  • A Praetorix HW tier proposed to fit
  • Personal contact and a date within 24 hours

New ticket — Praetorix AE

PRX-DEMO-••••

By submitting you agree to your data being processed in order to handle this request. More in our privacy notice.