What open-source AI does not automatically include
Open-source software comes with source code, free use and often an active community. What it does not automatically include: warranty, a contact person when something fails, tested security updates for how all parts work together, permissions from your directory service through to the specialist systems, a log nobody can alter, proof for data protection and the staff council, and training. These tasks fall to the operator, whether it handles them itself, hires a service provider or buys a ready-made platform.
What does open source include?
A great deal. Anyone may use, inspect, modify and pass on software under a recognised open-source licence, without a licence fee. Many building blocks for AI are mature and widely used: software that runs models on GPUs, interfaces, databases, sign-in services. Ready-made platforms build on them too; basebox uses vLLM and Keycloak, for example.
What open source is not: a product with commitments. The common licences Apache 2.0 and MIT provide the software “as is” and exclude warranty and liability as far as the law allows. Some projects have their own licences with conditions, such as Open WebUI with a clause on the brand name. So read the licence file before you deploy.
What is missing, and who provides it?
| Task | What the software provides | What the operator adds |
|---|---|---|
| Warranty and liability | usually excluded | own risk or a contract with a service provider |
| Security updates | updates per project, at its own pace | follow advisories, test how parts work together, install |
| Permissions | roles and groups in individual components | permissions from your directory service consistently through to the specialist systems |
| Connection | interfaces and extensions | build connections in which each person sees only what they are allowed to see |
| Log | settings per component | a log nobody can alter afterwards, with a retention period |
| AI model | free choice | test models, test switching, clarify the model’s licence |
| Protection against manipulation | depends on the project | checks for hidden instructions, permissions, human approval |
| Proof | none | DPIA, record of processing activities, TOMs, software bill of materials |
| Support | community, without commitment | contact person, cover, response time |
| Training | project documentation | AI literacy measures under Art. 4 of the AI Act |
Why does this only show up later?
A pilot usually runs with a few people, sample data and a committed colleague who knows everything. The missing parts only matter once real data is meant to go in: then data protection, information security and the staff council ask about permissions, logs and documents. And when that colleague goes on holiday, it becomes clear who else knows how things work. See the guide bus factor.
Which obligations stay with the operator?
- Data protection: Whoever uses the AI is the controller. They need a record of processing activities (Art. 30 GDPR), appropriate technical and organisational measures (Art. 32) and, for sensitive data, usually a data protection impact assessment (Art. 35). See DPIA for AI.
- AI literacy: Art. 4 of the AI Act requires measures for staff who work with AI, regardless of where the software comes from.
- Information security: Self-built AI belongs in your own ISMS and, where NIS2 obligations apply, in risk management. See ISO 27001 and NIS2 for self-built AI.
- Liability: Whoever builds an AI environment from open-source components may count as a manufacturer themselves. See Liability: open-source build or purchased platform?
The Cyber Resilience Act provides separate, lighter obligations for stewards of open-source projects (Art. 24). For the organisation that assembles and runs the components, this changes nothing about its own tasks.
When is building it yourself still the right choice?
When a team carries the tasks in the table permanently, for example in research or in a data centre with its own platform team. When very specific requirements cannot be met by any ready-made product. Or when the AI serves only a few specialists who need no connection to specialist systems. The guide TCO comparison compares the costs of both routes.
How does this work with basebox?
basebox includes the parts from the table: permissions from your directory service, connectors with each person’s own access, an audit log that cannot be altered, updates and support in the licence, and the data processing agreement, TOMs and agreement under § 203 StGB (German Criminal Code, professional secrecy) to download. Customers see the software bill of materials (SBOM) in their basebox; before purchase, we send it on request. Booked training sessions also cover AI literacy under the EU AI Act. If basebox becomes insolvent, the code will be published as open source so that customers can keep working.
Not legal advice: This guide gives an overview, as of 11 October 2026. Which obligations apply to your organisation depends on how you use AI and on your contracts.
Frequently asked questions
Is open-source software less secure than purchased software?
Not in principle. Open source code can be inspected, and many projects close vulnerabilities quickly. What matters is whether someone follows the advisories and installs updates. With open source, that is the operator's job.
Do we need a data processing agreement for self-hosted AI?
Not with the open-source projects, because they do not process any data on your behalf. If a service provider runs the environment for you or has access for maintenance, you need a data processing agreement with that provider under Art. 28 GDPR.
Does basebox itself use open source?
Yes. basebox is built on open-source components, such as vLLM for the AI model and Keycloak for sign-in. The components included are listed in the software bill of materials (SBOM). basebox itself is not open source.
Sources
- Open Source Initiative: The Open Source Definition ↗
- Apache License 2.0, sections 7 and 8 (disclaimer of warranty and limitation of liability) ↗
- MIT License (Open Source Initiative) ↗
- Open WebUI: licence with a clause on the brand name (documentation) ↗
- Regulation (EU) 2016/679 (GDPR), Art. 28, 30, 32 and 35 (EUR-Lex) ↗
- Regulation (EU) 2024/1689 (AI Act), Art. 4 (EUR-Lex) ↗
- Regulation (EU) 2024/2847 (Cyber Resilience Act), Art. 24 (EUR-Lex) ↗
This article was created with the help of AI, based on the sources listed. Responsible for the content: basebox GmbH. Not legal advice, no liability. Research as of 11 October 2026.