Security

Your data stays in the EU and in your hands.

Which certificates we hold, which laws and standards apply and what basebox does technically. With DPA, TOMs and templates for download.

How basebox protects your data.

DATA FLOW

Only what you enable goes out.

Vendor
basebox GmbHno access
Update →
Your protected area
Staff
Question ↓
Sign-inYour identity providerRoles, groups →
basebox
Update installed
AppsApp starts
ControlPermissions checked
ConnectionSearches shared content
AIAnswer created
Search ↓
Data
Request →
Internet
External modelblocked
Details
DirectionWhat is transmitted
Reachable from outside: a single access point.The interface and sign-in run over an encrypted connection (HTTPS). All other services communicate only internally.
From your server to basebox: nothing.No telemetry, no licence checks, no outbound update checks.
To you: software and models.They reach your environment with installation and updates, without internet as an offline package.
From you: only what you set up.Web search (Staan.ai, EU), an external language model (with the stated processing location), connected systems, an SMS service for two-factor sign-in, your SIEM or an external backup target.

ADDITIONAL APP WITH AN EXTERNAL MODEL (OPTIONAL)

The AI model in basebox runs with us. An external model is only available as a separate app.

Comparison of the two options
AI model in baseboxApp with an external model (optional)
Where it runsOn our servers in Germany or on your serverAt Google Vertex in Frankfurt, with models from Google or Anthropic, for example
Stores inputsNoNo training. Logging by Google for abuse detection, for at most 30 days.
Sub-processorsNone outside the EUGoogle, named in the list of sub-processors
US CLOUD ActDoes not applyApplies because Google is a US company
Who uses itEveryone, in the chat and in all appsOnly those who see the app: your administrator shares it with selected groups.
AvailableIn all operating modelsIn the cloud and in the starter environment

Other providers, such as OpenAI, can also be connected. We then point out the consequences for data protection and security.

Web search has no access to your internal data.

Details
RuleWhat this means for you
Separate assistants.Web content does not reach internal data. This is ensured by the architecture, not by the model's good behaviour.
Fixed outbound route.The connector reaches only the approved provider, without the person's identity or IP address.
Text only, off by default.No scripts. Web content is treated as data, not as commands. The person confirms before the first search.
Provider.Search runs via Staan.ai in the EU, in all operating models, only if you enable it. We have a DPA with Staan; training with the search queries is excluded.

PERMISSIONS

The AI may never do more than the person using it.

Sharing an app never shares data.

Rule for AI agents: An agent always acts on behalf of a person. It reaches only what has been shared with that person in basebox and what they are allowed to do in the target system themselves. Writing is off by default.

Connectors use the person's own access. There is no shared service account.

Three levels and questions on permissions

Permissions arise on three levels. Each is managed where it already lives today.

Which user directory can we connect?
Any that speaks LDAP, OIDC or SAML, for example Active Directory, Entra ID or WSO2. Sign-in runs via Keycloak in basebox; basebox never sees passwords. A second factor is possible.
Who may use which app?
Whoever built the app shares it: with groups from your user directory or with individual people, for use only or also for editing.
Who sees which data?
Each person sees only what they are allowed to see in the respective system anyway. They sign in there via the connection with their own access; basebox uses no shared account.
How does a person sign in to a system?
Via OAuth with their own account in the respective system. basebox receives an authorisation for this, not a password. There is no shared service account.
Where do the groups come from?
From your user directory, for example Active Directory. If something changes there, it also applies in basebox.
What happens when someone shares an app?
What is shared is the task, not the access. Anyone using a shared app still sees only their own data in it.
Who may use which AI model?
Your administrator enables models, like connections to systems. In an app, only an enabled model can be selected.
What may an automated workflow do?
Only what the person starting it is allowed to do. The Workflow Builder, too, can only include connections and models that your administrator has enabled.
Can we set binding rules for everyone?
Yes. Your administrator sets usage policies and system defaults centrally. They apply to all apps in your organisation.

TRACEABILITY

Every answer shows where it comes from.

Clicking the source opens the passage in the document.

Details
  1. What it is based on: sourcesEvery statement from documents carries a source marker. One click shows the document, the quoted passage and how faithful the answer is to the source. The original can be downloaded.
  2. How it came about: model and line of reasoningWhich model in which version answered is documented. The model's line of reasoning can be expanded.
  3. Who decides: the professionalThey check at the source and are responsible for the use. Write actions by agents only after approval.
  4. What was done: audit logA table for administrators: who, when, which assistant, which connector. Export as CSV.
Boundary: Why a model calculates internally the way it does depends on the model and its training. Even a displayed line of reasoning can be wrong. The yardstick is therefore the source, not the model's reasoning.

LOG, DELETION, BACKUP

Actions are logged. Content only if you switch it on.

The audit log cannot be changed and can be exported for your SIEM.

Full security description ↗
Details
TopicHow basebox handles it
The audit log records who did what and when:Sign-ins, permission changes, configuration, chats, workflows. Only identifiers, no names or e-mail addresses. Entries cannot be changed or deleted individually. Retention 6 months, adjustable from 3 to 36 months. Export as CSV, on request for your SIEM.
The content of questions and answers is excluded,as long as your organisation does not switch on detailed logging. If your administrator switches it on, the questions are logged as well.
System logs are deleted after 7 days,unless you set otherwise. Questions, answers and documents are not included.
Deletion happens on your side:People delete their own chats and documents; administrators delete everything else.
With the server, the backup is with you,with your tools. basebox has no copy of it and no access to it.

On the safe side legally.

GDPR

basebox fulfils the obligations of a processor.

Processing only on your instructions, in Germany or on your server.

Details

In the cloud and when basebox operates the server, we process data on your behalf. With a server we do not operate, we process your data only if you give us access for support.

  • Data processing agreement (DPA) under Art. 28 GDPR: part of every contract.
  • Technical and organisational measures (TOMs) under Art. 32 GDPR, for download.
  • Processing only on your instructions, in Germany or on your server. Web search runs in the EU. Your data does not train any model.
  • Patient, citizen and social data may be processed, on the server and in the cloud.
  • We use sub-processors only if they are named in the contract. We inform you in advance of planned changes, and you can object.
  • We report breaches of the protection of your data to you without undue delay (Art. 33(2) GDPR).
  • We support you with requests from data subjects and with your impact assessment.
  • After the contract ends, we delete or hand over all data.
  • You can reach our data protection officer at datenschutz@basebox.ai.

The version agreed in your contract is authoritative.

For professional secrecy holders, an agreement under Section 203 of the German Criminal Code (StGB) applies.

It is part of your licence agreement.

Details

Doctors, psychotherapists and other professional secrecy holders may only involve service providers if these are bound to confidentiality. The agreement under Section 203 of the German Criminal Code (StGB) is an annex to the licence agreement. It is included if you or your staff are professional secrecy holders.

Church-run organisations work with basebox under the KDG and DSG-EKD.

For Caritas, Diakonie and church-run hospitals.

Details

Catholic institutions, including Caritas hospitals, work under the Catholic Church data protection act (KDG) and its implementing ordinance (KDG-DVO). Protestant institutions and the Diakonie work under the data protection act of the Protestant Church in Germany (DSG-EKD). In both acts, the obligations of a processor largely correspond to Art. 28 GDPR.

  • Protestant institutions: the DPA under Art. 28 GDPR is sufficient (Section 30 DSG-EKD).
  • Catholic institutions: in addition to the DPA, we agree on the application of the KDG (Section 22 KDG-DVO). On request, we send you the signed DPA and supplement, as Section 29 KDG requires.
  • Health and personnel data fall under data protection class III of the KDG-DVO. basebox transmits data encrypted; a second factor at sign-in is possible.
  • Our ISO 27001 certificate serves as your evidence under Section 9(7) KDG-DVO.
  • After the contract ends, we delete or hand over all data.

FADP, Art. 321 SCC

Swiss hospitals and public authorities work with basebox under Swiss law.

Cloud and server, with contract annexes on data protection and professional secrecy.

Details

For customers in Switzerland, two annexes supplement the licence agreement and the DPA: a supplement under the Federal Act on Data Protection (FADP, DSG) and a confidentiality agreement under the Swiss Criminal Code. Cantonal and municipal bodies are subject to the data protection act of their canton; the DPA supplement takes this into account.

  • Cloud: currently runs in Germany, a country with adequate data protection under Annex 1 of the Data Protection Ordinance (DPO, DSV).
  • Swiss data centre: in progress for the cloud, by 2027 at the latest.
  • Server: our hardware partner FAST LTA delivers it to Switzerland.
  • Processing on behalf: supplement to the DPA under Art. 9 FADP (DSG), including notification of data security breaches under Art. 24 FADP.
  • Confidentiality: basebox and all employees involved are bound in writing as auxiliary persons, for professional secrecy (Art. 321 SCC) and official secrecy (Art. 320 SCC).

DATA PROTECTION IMPACT ASSESSMENT

You get a template for your impact assessment.

The details about basebox are already filled in.

Details

For AI with health or social data, a data protection impact assessment (DPIA) is usually required. You prepare it and are responsible for it as the controller. Our templates are structured according to Art. 35 GDPR and the Standard Data Protection Model: the details about basebox are pre-filled, examples are marked, your parts are empty. The same is available for your record of processing activities.

You do not need to anonymise your data.

The AI works with the complete documents, within your boundaries.

Details

Anyone using an AI outside their own organisation often removes names and identifiers first. That offers only limited protection:

StatementReason
Pseudonymised data remains personal data.Under the GDPR, they are still personal data, so the obligations remain.
Free text cannot be reliably anonymised.Doctors' letters, applications and minutes contain clues to people even where no name appears.
What is removed is missing for the AI.Without context, the answers get worse.

With basebox, the data stays within your boundaries. The AI works with the complete documents; nobody redacts anything beforehand.

WORKS AND STAFF COUNCIL

Logging is for security, not for performance monitoring.

Roles come from your user directory; no new profiles are created.

Model works agreement (Word)Template and process in the guide →
Details
TopicHow basebox handles it
Purpose of logging.The logs serve the security of the system, not the assessment of performance or behaviour.
No new profiles.Roles and permissions come from your existing user directory; there is no second data store.
Control over automation.Agents act only with the permissions of the person using them.
Qualification.Booked training sessions also cover AI literacy under the EU AI Act. It can also be booked separately.

Certificates and evidence.

basebox is certified under ISO 27001.

This applies to the management system of basebox GmbH.

Details

The information security management system of basebox GmbH is certified under ISO/IEC 27001. You can download the certificate and scope.

BSI C5

The C5 Type 1 attestation is in progress.

The pre-audit and attestation for our cloud are scheduled.

Details

C5 is the BSI's criteria catalogue for cloud services (BSI is Germany's Federal Office for Information Security). Under Section 393 SGB V, cloud providers that process health data need a current C5 attestation. The pre-audit and attestation for our cloud are scheduled. With the server, C5 is not required; your requirements as the operator apply there.

basebox is implementing the Cyber Resilience Act.

The software bill of materials and the reporting channel are in place; the remaining obligations follow by December 2027.

Details

The Cyber Resilience Act (CRA) applies to software placed on the market in the EU. The reporting obligations for exploited vulnerabilities have applied since 11 September 2026, the other obligations from 11 December 2027.

Done
  • Software bill of materials (SBOM): generated automatically and updated with every release. Customers see it in full in their basebox; before purchase, we send it on request.
  • Reporting channel for vulnerabilities with security.txt, see below.
In progress
  • Processes for vulnerabilities and security updates throughout the entire support period: in progress.
  • Technical documentation and EU declaration of conformity: in progress, before the deadline.

basebox is not a medical device and therefore does not need ISO 13485.

For the platform, information security is what counts.

Details

ISO 13485 governs the quality management of medical device manufacturers. A medical device is an application with a medical intended purpose, such as a diagnostic suggestion for a patient. basebox is a platform for general tasks such as drafting letters or searching guidelines. For it, information security is what counts: ISO 27001 and C5.

If you build a medical device yourself and use models from basebox in it, you are the manufacturer. For your technical documentation, we provide the information of a supplier: model register with version and licence, software bill of materials and ISO 27001 certificate.

ISO 27001 is the basis for TISAX and SOC 2.

basebox does not have a TISAX label.

Details

TISAX is the automotive industry standard and builds on ISO 27001. basebox does not have a TISAX label. If your company needs one, talk to us. SOC 2 is a US audit standard. In Europe, ISO 27001, supplemented by C5 for cloud services, is the usual evidence.

For your obligations as an operator.

basebox provides the technology and the evidence for this.

You get the evidence for NIS2, KRITIS and the B3S for hospitals.

Separate environment and immediate reporting of incidents.

Details

Hospitals and many public sector organisations must secure their IT under NIS2 and the KRITIS requirements (German rules for critical infrastructure), hospitals under the industry standard B3S for hospitals. This also applies to AI.

  • Separate environment: in the cloud, your own environment separated from other customers; with the server, the hardware on your premises, without internet if you wish.
  • Evidence for your review: ISO 27001 certificate, TOMs, description of architecture and data flows. The C5 Type 1 attestation is in progress.
  • Reporting obligations: we report security incidents to you immediately. The reporting channel and contact persons are set out in the contract so that you can meet your own deadlines.
  • Supply chain: development, support and operation in Germany, software bill of materials (SBOM) for every release.
  • Logs: audit log, on request as an export for your SIEM.
  • Professional secrecy: agreement under Section 203 of the German Criminal Code (StGB).
  • B3S for hospitals (industry-specific security standard): we are currently mapping our measures to the B3S requirements. You can get it on request.

BSI IT-GRUNDSCHUTZ

You get the documentation for your ISMS under IT-Grundschutz.

What is certified is your information domain, not a single product.

Details

IT-Grundschutz applies to the operator's information domain, not to a single product. A product therefore cannot be certified under IT-Grundschutz. For your modelling, we provide architecture, data flows, interfaces, roles and logs.

For the EU AI Act, you get training, logs and a model register.

basebox is a platform, not a high-risk system.

Details

basebox is a platform, not a high-risk system. Which obligations you have depends on how it is used.

  • AI literacy (Art. 4, mandatory since February 2025, since July 2026 as an obligation to promote it): we train your staff, virtually or on site. Or you train them yourself, with our slides and templates.
  • Model register: developer, version and licence of each model are documented.
  • Logs: the audit log records who used which assistant and which connector.
  • Human oversight: write actions by agents require approval.

Documents.

Report vulnerabilities to security@basebox.ai.

We confirm receipt within two working days.

Details

We confirm receipt within two working days, keep you informed of progress and, if you wish, credit you by name once the vulnerability is closed. Our contact details are machine-readable in the security.txt.

STARTER ENVIRONMENT

New accounts start in a shared environment.

At Hetzner in Nuremberg, with AI models from Google Vertex in Frankfurt.

Details

Anyone who registers on the website starts in our starter environment at Hetzner in Nuremberg. All customers use a shared environment there and are separated from each other as tenants. There is no separate environment or separate user directory per customer there. In the starter environment, the AI models come from Google Vertex in Frankfurt: Google does not use your inputs for training and logs them for abuse detection for at most 30 days. Backups are made daily and kept for up to 7 days. You decide whether you process personal data there. For health data and other special categories under Art. 9 GDPR, we recommend your own environment in the cloud or the server on your premises.

Frequently asked questions

All questions →

Cookie settings

This website sets no cookies for analytics or advertising and embeds no tools from other providers. Your browser only stores what is needed to use the site. Privacy policy