Data protection and security

For data we process on your behalf, we are the processor and you are the controller. Before we process real data, we conclude a data processing agreement (DPA) with you under Art. 28 GDPR. It ties the data to the purpose of the engagement. Our standard is Variant B. We conclude the contracts with the AI providers ourselves. The providers are then our sub-processors, and you approve them in the DPA. In Variant A, you engage the AI providers yourself and instruct us to work through your accounts. Large customers may keep their own contracts.

Our services run on a virtual server at Contabo GmbH. It is located in the Lauterbourg data centre in France, in other words in the EU. Each application runs in its own container with its own data directory. Where the system allows it, customer data stays on the customer's servers. If a product stores data with us, the data sheet states the location, the period and the purpose. If a product requires processing in the EU or the EEA, it processes real data only through Azure OpenAI from Microsoft or through open models at Cortecs. At Microsoft, the resource is located in Sweden, and depending on the deployment, Microsoft processes within the EU data zone or only in Sweden. At Cortecs we approve only data centres in the EEA, including in Iceland, which is part of the EEA but not of the EU. Not every product has been switched over yet. Where the code still calls OpenAI directly, we change it before the product processes real data. Other products use OpenAI or Anthropic, with processing in the USA as well. Your DPA states which provider it is.

These seven companies are on our list of sub-processors. Which of them act for you depends on the product and is stated in Annex 3 of your DPA. If you engage the AI providers yourself (Variant A), Microsoft, OpenAI, Anthropic, Cortecs and Mistral do not apply to you.

We do not train any models with your data. The API terms of Microsoft, OpenAI, Anthropic and Cortecs do not allow the providers to use inputs and outputs for training. Microsoft stores inputs and outputs for up to 30 days in the EU for abuse monitoring. They are reviewed only on suspicion, by staff in the EEA. We cannot switch this storage off: on 2 November 2026 Microsoft declined our request to do so (Modified Abuse Monitoring), because the feature is open only to managed customers and partners of Microsoft. OpenAI stores data for up to 30 days for abuse detection. Anthropic deletes the API's inputs and outputs within 30 days. Cortecs keeps inputs and outputs only in memory and stores request metadata for up to 12 months. We approve only data centres that, according to Cortecs, do not store inputs and outputs.

A selection of the measures we have implemented. We send the complete technical and organisational measures (TOMs) under Art. 32 GDPR, with the status of each measure, on request.

• Responsibility

  Philipp Gurndin provides the services personally and is responsible for data protection and security. Other people work only with your prior written consent and after a written commitment to confidentiality.

• Confidentiality

  We are bound to confidentiality, with no time limit. The basis is the project contract and the non-disclosure agreement.

• Sign-in

  Sign-in to our web interfaces runs through Invariance's sign-in service, with personal accounts and a second factor. Roles apply per application.

• Separation of customers

  Each customer has a separate, delimited area. Each application runs in its own container with its own data directory.

• Dedicated access at the AI providers

  For processing on your behalf, we use language models only through API accounts with a data processing agreement. For each customer we create a separate key there, at Microsoft a separate resource, at OpenAI a separate project.

• Access to the server

  Only Philipp Gurndin administers the server. SSH works only with a key, passwords are switched off, and access is blocked after failed attempts.

• Country filter

  Administration, internal web interfaces, SSH and sign-in to the mail system respond only from approved countries. Only websites and products are publicly reachable.

• Access to your systems

  We access your systems only by the route you specify, for example through your VPN, and with personal accounts. Your administrator assigns the rights.

• Transmission

  All web services and interfaces run over HTTPS with TLS. We transmit mail with TLS.

• Keys and passwords

  The keys and passwords of our services are stored encrypted in the server's configuration or in our password vault. They are not in the source code, not in logs and not in extensions on workstations.

• Logs

  We log access to the web server without headers, on a rolling basis, up to a maximum of 30 MB. Our applications do not write customer inputs to logs.

• Development and testing

  We develop and test with fictitious data or data you have approved. We use ChatGPT and Claude subscriptions only for these limited tests. Source code repositories contain no customer data.

• Backup

  We back up daily, including the sign-in service, mail and customer area. We have tested restoration, most recently on 24 October 2026.

• Updates

  The operating system receives security updates automatically. We update the platform and applications regularly, most recently on 24 October 2026.

• Monitoring

  We check the services automatically, with an alarm to Philipp Gurndin. A central error log runs on our own server.

• Outage

  If an Invariance service fails, your system remains usable to the extent the data sheet does not say otherwise.

• Deletion at the end of the contract

  After the contract ends or on request, we delete data, copies and access, and confirm the deletion in writing. Data in backups expires with their rotation and stays blocked until then.

An incident is any event in which a customer's data could be lost, altered or made accessible to unauthorised persons, and any disruption that affects the security of a service. Even a suspicion counts. We report to you without undue delay any incident that may affect your data, at the latest 12 hours after we become aware of it. The period begins when Philipp Gurndin learns of it. If a sub-processor reports an incident to us, the same period applies from the time we receive its report. The report goes by email to the person authorised to give instructions on your side and to your technical contact, and, if the situation requires it, also by phone. It states, as far as known, the information required by Art. 33(3) GDPR. We document every incident, even if the suspicion is not confirmed. You make the notifications to the authorities as the controller. This leaves you time for your own deadlines. We supply you with the technical details for this.

1. Immediately

   We contain the incident before it has been fully clarified. We switch off affected services or take them off the network, block access, replace keys and passwords and stop data transfers to providers. We preserve traces before anything is deleted or rebuilt.

2. 12 hours

   At the latest by then, we report the incident to you, counted from when we become aware of it.

3. 24 hours

   Your deadline for the early warning to the competent CSIRT under NIS2, in Italy CSIRT Italia, if you fall under NIS2.

4. 72 hours

   Your deadline for the notification to the supervisory authority under Art. 33(1) GDPR, in Italy the Garante.

5. After completion

   You receive a written report: sequence of events, cause, consequences, measures and what changes permanently.

Depending on the product, we prepare these documents for you before the start.

• Data processing agreement (DPA)

  Under Art. 28 GDPR, with three annexes: the technical and organisational measures, the contact points and the list of sub-processors. The standard is Variant B.

• Data sheet

  One sheet per product and customer for your software and supplier inventory under NIS2. It states which data the product reads, sends and stores.

• Data protection impact assessment (DPIA), technical part

  The technical part of a data protection impact assessment under Art. 35 GDPR. You add the remaining part.

• Zero data retention overview

  Which AI provider stores inputs and outputs, with evidence.

• AI guidance

  Suggestions for four steps when using a product with AI at your organisation: informing patients or clients, instructions to your employees, staff training, approval before routine operation. You adopt or amend the texts.

• Non-disclosure agreement

  The agreement on confidentiality. It can also be concluded before the first meeting with real data.

• Project contract

  The contract for the project. A separate section governs professional secrecy, data protection and data security. Before production use, you can commission a security test by a third party at your own cost.

Does data go to AI providers or other services? To which ones, in which countries?

For processing on your behalf, only to the sub-processors in Annex 3 of your DPA or to providers with whom you have a contract yourself. If a product requires processing in the EU or the EEA, it processes real data only through Azure OpenAI from Microsoft in the EU, with the resource in Sweden, or through open models at Cortecs, only at data centres in the EEA that we approve, including in Iceland. Where the code still calls OpenAI directly, we change it beforehand. Other products use OpenAI or Anthropic, with processing in the USA as well. Your DPA states which provider it is. Our own services run on a server at Contabo in France.

Are inputs, results, errors or usage logged?

Customers' inputs and results do not end up in logs. We log errors without content. We log access to the web server without headers and on a rolling basis.

Is data used to train or improve AI models?

Not through the APIs we use on your behalf. We do not train any models with your data, and the API terms of Microsoft, OpenAI, Anthropic and Cortecs do not allow the providers to use inputs and outputs for training.

Is there a data processing agreement that excludes other purposes?

Yes. Before we process real data, we conclude a DPA with you under Art. 28 GDPR, which ties the data to the purpose of the engagement. We have our own contracts with Microsoft, OpenAI, Anthropic, Cortecs, Mistral, Contabo and Brevo.

Can data transmission be limited or the software be switched off at short notice?

Yes. You can block extensions by policy and block the addresses of our services at your firewall. We can switch off immediately any service that we operate.

Philipp Gurndin gurndin@invariance.co +39 350 0404 001 PEC gurndin@pec.it