For CISOs, CIOs and DPOs of SMEs and mid-market

Your realm.Your rules.Even with AI.

MaelGard analyses every prompt sent to generative AI on the device and applies your company rules before data leaves your realm. Depending on the rule, the prompt goes through, is masked, waits for its author's decision or is blocked.

  • 104 AI services recognised, and unknown AI detected automatically
  • 8 ready-to-use detection models, or custom-built ones
  • 100% on-device analysis: no prompt is sent to MaelGard or any third party for analysis
  • Hosted in France (Scaleway, Paris) or installed on your premises, including air-gapped networks
Your realm · HR workstation3 sensitive values

Can you rephrase this message: “Hello Mark Dawson, following your email mark.dawson@acme-corp.example, your bonus will be paid to IBAN GB29 NWBK 6016 1331 9268 19.”

MaelGard checkpoint
Out of realm · AI serviceSent, masked

Can you rephrase this message: “Hello [NAME_1], following your email [EMAIL_1], your bonus will be paid to IBAN [IBAN_1].”

Fictional example, simulated in your browser without any network request.

Catalogue

104 AI services recognised

Web channel
  • ChatGPT
  • Claude
  • Gemini
  • Microsoft Copilot
  • Grok
  • DeepSeek
  • Le Chat (Mistral)
  • Perplexity
  • Poe
  • HuggingChat
  • Meta AI
  • Character.AI
  • You.com
  • Venice
  • Qwen Chat
  • Qianwen (Tongyi)
  • Kimi
  • ChatGLM (Zhipu Qingyan)
  • Z.ai Chat (GLM)
  • Doubao
  • ERNIE Bot (Wenxin Yiyan)
  • Tencent Yuanbao
  • MiniMax Agent (Hailuo)
  • Microsoft 365 Copilot
  • NotebookLM (Gemini Notebook)
  • Google AI Studio
  • Duck.ai
  • Opera Aria
  • Pi
  • Kagi Assistant
  • Felo
  • Genspark
  • Arena (LMArena)
  • Monica
  • Merlin
  • MaxAI
  • Mammouth AI
  • Abacus ChatLLM
  • T3 Chat
  • Lumo
  • Notion AI
  • DeepL Translator / Write
  • QuillBot
  • GitHub Copilot Chat (github.com)
  • v0
  • bolt.new
  • Midjourney
  • Suno
  • Udio
  • OpenRouter API
  • ElevenLabs API
API channel
  • DeepL Translator / Write
  • OpenAI API
  • Claude API
  • Gemini API
  • xAI API
  • DeepSeek API
  • Mistral AI API
  • Perplexity API
  • Azure OpenAI / Microsoft Foundry
  • Amazon Bedrock
  • Vertex AI (Gemini Enterprise Agent Platform)
  • Cohere API
  • Groq API
  • Together AI API
  • Fireworks AI API
  • OpenRouter API
  • Hugging Face Inference Providers
  • Replicate API
  • Cerebras Inference API
  • SambaNova Cloud API
  • DeepInfra API
  • Nebius Token Factory
  • Novita AI API
  • Hyperbolic API
  • NVIDIA API Catalog (NIM)
  • Alibaba Model Studio (Qwen / DashScope)
  • Kimi API (Moonshot AI)
  • Z.ai / BigModel API (GLM)
  • MiniMax API
  • Baidu Qianfan (ERNIE)
  • Volcengine Ark / BytePlus ModelArk (Doubao, Seed)
  • AI21 Studio API
  • Reka API
  • Writer API (Palmyra)
  • Inflection AI API (Pi)
  • LightOn Paradigm
  • Scaleway Generative APIs
  • OVHcloud AI Endpoints
  • IBM watsonx.ai
  • Cloudflare AI Gateway
  • Vercel AI Gateway
  • Baseten Model APIs
  • Featherless AI API
  • Nscale Serverless Inference
  • Tencent Hunyuan API
  • StepFun API
  • Upstage API (Solar)
  • Stability AI API
  • ElevenLabs API
  • Black Forest Labs API (FLUX)
  • Runway API
  • Luma Dream Machine API
  • Ideogram API
  • fal.ai API
  • Voyage AI API
  • Jina AI API

AI outside the catalogue is detected automatically from the structure of requests. The API channel must be enabled by an administrator. Names are descriptive and trademarks belong to their owners.

How it works

A checkpoint between typing and sending

Your teams keep their AI tools. MaelGard steps in when a request leaves the device and applies your company policy.

  • It goes through

    Nothing sensitive, or the rule allows it; the event is logged.

  • It is masked

    Sensitive values are replaced by labels such as [EMAIL_1].

  • It is blocked

    The rule forbids this sharing; nothing is sent.

  • The person decides

    A window shows what was detected and the allowed choices.

Detection models

8 ready-to-use detection models

Each model spots one family of sensitive data. Model by model, you choose what goes through, what is masked, what is blocked and what requires a decision.

  • Personal data

    Names, emails, phone numbers, addresses, dates of birth, ID numbers, IBANs, bank cards, IP addresses, identifiers, secrets

    Plan : Solo, Essential, Complete

  • Finance

    Amounts, percentages, IBANs, BICs, ISINs, ticker symbols, bank cards

    Plan : Complete

  • Human resources

    Employee numbers, job titles, departments, pay, contracts, grades, leave, appraisals, disciplinary actions, terminations

    Plan : Complete

  • Technical

    Credentials and passwords, IP and MAC addresses, host names, URLs, file paths, ports, hashes, software, versions, CVEs

    Plan : Complete

  • Geolocation

    Addresses, postcodes, cities, regions, countries, GPS coordinates, landmarks

    Plan : Complete

  • Telecom

    IMEI, IMSI and ICCID numbers, phone numbers, carriers, equipment, IP and MAC addresses, volumes and data rates

    Plan : Complete

  • Does your business have its own sensitive data?

    We build custom models: product references, customer codes, internal nomenclatures…

    Talk to us

Sovereignty by design

What happens in your realm stays there.

Requests are analysed on the device. Prompts and texts never leave your environment: only decision metadata (category detected, AI service, rule applied) is sent up to the console.

Two options, available today:

  • Online

    Console hosted in France, in Scaleway's Paris data centres.

  • On your premises

    Server you can install on site, including on an air-gapped network.

Use cases

Five ordinary work situations

  • 01

    Human resources

    To rephrase a letter or prepare an interview, people paste text into the AI that often holds an employee's personal data.

    Detected: names, email addresses, contact details.

  • 02

    Finance and administration

    A payment reminder to check or a billing thread to summarise often contains an IBAN or other bank details.

    Detected: IBANs and bank details.

  • 03

    Management and technical teams

    A management memo or a code snippet pasted into an AI can contain a password, an access key or an internal figure.

    Detected: passwords and access keys, IP addresses and server names, project codes, “Confidential” markings, internal figures and prices.

  • 04

    Health and social care

    AI saves time on a medical report to summarise or a patient letter to rephrase. These texts contain medical information about identifiable people.

    Detected: diagnoses, treatments, dosages, test results, record numbers.

  • 05

    Industry and R&D

    A specification to write or a supplier comparison often contains your process parameters and your prices.

    Detected: process parameters, materials, part numbers, project codes, prices.

Comparison

MaelGard compared with other approaches

A comparison by category of solution, to govern AI use across the whole company.

Comparison of solution categories to govern generative AI use
CriterionFree anonymisation toolsOffice suite DLPAI security platformsMaelGard
Company policy and consoleno : Individual useYesYesYes
Analysis on the deviceYesYespartly : Often in the cloudyes : Embedded models
Third-party AI coveredPartlypartly : Depends on licencePartlypartly : Web and API catalogue services, according to enabled channels and rules
Group Policy or Intune deployment, auditNoYesYesYes
Hosting in France or on-premisespartly : Variespartly : Vendor cloudpartly : Rarelyyes : On-premises (including air-gapped) or online in France
Compliance

Document AI use, decision by decision

The DPO and the CISO get the rules applied, the decisions taken and their reasons, without prompt content.

  • Rules by category

    Define, category by category, what goes through, what is masked, what is blocked and what requires a decision.

  • Audit log

    Keep decision metadata, including the reason for unprotected sending.

  • On your premises or in France

    The online service is hosted in France, by Scaleway (Paris region). The server can also be installed on your premises, including on an isolated network.

  • No usage telemetry

    No content or AI usage is sent to us. Only a technical licence report (active devices, models, version) may be, and it can be turned off.

GDPR and the AI Act: this tracking data (categories detected, AI services used, decisions) feeds your record of processing; the next section shows how MaelGard shapes it for your audits.

Transparency

In place, coming, and known limits.

In place

  • 8 detection models: personal data, finance, HR, health, technical, trade secrets, geolocation, telecom
  • Custom models
  • Automatic detection of AI outside the catalogue, from request analysis
  • Usage policy displayed on the device
  • Agents for Windows (x64, ARM64), Linux (x86_64, aarch64, Flatpak) and Android
  • Sealed audit log
  • Mutual TLS between agent and server
  • Group Policy and Intune deployment
  • Hosting in France (Scaleway, Paris) or on-premises

Coming

  • Self-service opens on 1 January 2027
  • Audit log timestamped and signed by Certigna (RFC 3161) from 1 January 2027
  • Open-source release of the signing and verification modules on GitHub
  • macOS version under consideration

Known limits

  • AI services may change their protocols without notice: we update the catalogue continuously, and automatic detection takes over in the meantime.
  • The API channel must be enabled by the administrator.
  • Detection depends on channel, browser and request format.

Details on the Security page

Pricing

Simple pricing, per device

Prices exclude VAT. Self-service opens on 1 January 2027; until then, we deploy by appointment.

Comparison of the Solo, Essential, Complete and Custom plans
PlanSoloEssentialComplete RecommendedCustom
PriceFree€2 excl. VAT / device / month€3 excl. VAT / device / monthCustom quote
Devices1Up to 1,000Up to 1,000Over 1,000
ModelsPersonal dataPersonal dataAll 8 modelsAll 8 models + custom models
Choice of actions (allow, mask, ask, block)YesYesYesYes
Admin consoleNoYesYesYes
Audit logNoYesYesYes
Group Policy and Intune deploymentNoYesYesYes
Get notified of the launchGet notified of the launchGet notified of the launchContact us
FAQ

Frequently asked questions

How do I stop employees from pasting sensitive data into ChatGPT?

Banning ChatGPT moves the problem to hidden usage. MaelGard steps in at the moment of sending: the agent analyses the text on the device and, if it spots sensitive data, lets it through, masks it, blocks it or asks the person to decide, according to the company policy deployed through Group Policy or Intune.

How is it different from a traditional DLP such as Microsoft Purview, which already protects Copilot?

A traditional DLP such as Purview mainly protects files, email and Microsoft 365 data, based on rules; protecting Copilot often requires premium licences. MaelGard analyses text sent to AI services, both those in its catalogue and those it detects automatically, without changing your office licence. Detection runs on the device and the person decides when sending. The two approaches complement each other.

Does MaelGard help with GDPR and the AI Act?

MaelGard reduces the risk of personal data being transferred to third-party AI services and produces usable evidence: every decision is written to a sealed audit log, where any change or deletion is detectable. From 1 January 2027, each entry will also be timestamped and signed by Certigna, a French trust service provider, under the RFC 3161 standard. It contributes to your approach without replacing it: the record of processing and impact assessment remain yours.

How can I check that the audit log has not been tampered with?

The audit log is sealed: any change to or deletion of an entry is detectable. From 1 January 2027, each entry will be timestamped and signed by Certigna under the RFC 3161 standard, and the signing and verification modules will be published as open source: you, your auditor or your DPO will be able to check every signature with your own tools, without having to trust us.

What is Shadow AI?

It is the use of AI tools by employees without approval or visibility from IT: personal accounts, extensions, unlisted applications. The MaelGard dashboard shows which AI services are used and which categories of sensitive data are detected, from metadata only, to limit data leaks.

What if my teams use an AI that is not in your catalogue?

MaelGard recognises 104 AI services. For the others, the agent automatically detects requests that look like exchanges with a generative AI, based on their structure. You are alerted, and your rules also apply to tools you did not know about yet.

How does MaelGard step in before sending?

The agent spots requests to AI services. Depending on enabled channels and rules, their content is analysed on the device before sending. Web and API rules are distinct; acceptance should cover the actual browser and provider. Other traffic is not inspected.

Is MaelGard ready for production?

Yes. The Windows and Linux agents install through Group Policy, Intune or Flatpak, the Android agent by installing its app, and we support every deployment directly. Self-service opens on 1 January 2027; until then, we deploy by appointment.

Can I try MaelGard for free?

Yes, from 1 January 2027. The Solo plan, free for one device, protects personal data and lets you choose what goes through, what is masked or blocked. It includes neither the console nor the audit log.

Get notified of the launch

Can you detect data specific to our business?

Yes. On top of the 8 ready-to-use models, we build custom models (product references, customer codes, internal nomenclatures). They are offered on quote.

Does it cover health data?

The Health model spots diagnoses, treatments and test results before they leave for an AI. Analysis stays on the device: no health data is sent to MaelGard.

Why is MaelGard so affordable?

The low price comes from the architecture, which is cheap to run. Analysis runs on the device, so no cloud compute is re-billed for each prompt. The server only receives metadata and hosting stays light. Deployment uses standard tools (MSI, Group Policy, Intune), with no proxy or gateway to add.

Can unprotected sending be allowed?

Yes, if the company policy allows it. The person then gives a reason, and the decision is kept in the audit log, without the sensitive value.

Which data is stored, and where?

On the device: configuration and an encrypted local log. On your company server: metadata (category detected, AI service, rule applied, decision, date, device or pseudonym), never prompt text. The online service is hosted in France, by Scaleway (Paris region); the server can also be installed on your premises, including on an isolated network.

Does MaelGard work on Mac?

Not yet. MaelGard currently protects Windows (x64, ARM64) and Linux (x86_64, aarch64) devices, as well as Android devices. A macOS version is being considered; leave us your email to be notified.

Notify me

Does MaelGard install a kernel driver?

On Windows, the MSI installs the drivers needed to analyse requests. On Linux, the host service uses nftables and system mechanisms without an additional MaelGard kernel driver. Procedures and limitations differ by operating system.

What are MaelGard's limits?

MaelGard greatly reduces the risk of accidental sending, without claiming to stop everything: a local administrator can bypass the agent, and applications with pinned certificates cannot be analysed. The console flags these situations so you stay in control.

Open AI to your teams. Keep your data in your realm.

Let's look at your context together in 30 minutes: a demo on a device and your deployment questions.