TÜV Association and TÜV AI.Lab Present Concept for Testing Frontier AI

Berlin, 2 November 2026 – On the occasion of the German government's Digital Summit, the TÜV Association and TÜV AI.Lab are presenting a proposal to establish an independent testing and oversight structure for frontier AI and autonomous AI agents. Germany and Europe must be able to assess the safety of these AI systems independently, regardless of where they were developed. In the view of the TÜV organisations, AI models should be tested before they are placed on the market. In future, the German AI Security Institute (AISI) should play a central role in cooperation with independent testing organisations. The TÜV Association and TÜV AI.Lab explain their proposal in a joint white paper.

"Europe must, of course, aim to develop its own powerful AI models. But we must prepare for the fact that a significant proportion of the AI used in Europe will continue to come from the US and other regions of the world," says Dr Joachim Bühler, Managing Director of the TÜV Association. "And when assessing their safety, we must not rely solely on manufacturers' claims, particularly when critical infrastructure and national security are concerned. We must be able to assess for ourselves what risks are associated with their use and what protective measures are necessary. Digital sovereignty therefore means not only developing our own technologies, but also being able to independently test and control technologies developed elsewhere."

Establish proven testing structures for AI

Independent testing structures have long been established in Europe for technologies that can endanger people or the environment – for example, vehicles, industrial installations and medical devices. TÜV AI.Lab and the TÜV Association are calling for independent testing structures to be established for frontier AI based on this model. This is because AI is becoming more powerful and is increasingly being integrated into digital and physical systems. Testing should begin during development and accompany a model throughout its entire lifecycle. Its capabilities, the effectiveness of its safeguards and its behaviour under realistic operating conditions should be examined. To this end, independent testers need access to the models, their training processes and safety-relevant information about the development process. Binding requirements for testing methods, qualifications and the independence of the bodies involved are also needed.

"Independent testers must be able to scrutinise developers' safety claims," says Franziska Weindauer, Managing Director of TÜV AI.Lab. "To do so, they need access to the models and insight into training and safety measures. They must be able to independently assess whether existing risks have been sufficiently mitigated. And when AI controls a machine, machinery safety experts must also be involved in testing. They can assess how AI affects existing safety mechanisms."

Combining government responsibility with independent testing

Under the proposal, AISI should gradually be developed into an independent institution with a clear mandate and sufficient resources. It should build up its own expertise in assessing the safety of frontier AI while also involving independent testing organisations with specialist knowledge of industry, mobility and energy supply. This would enable the institute to become operational quickly.
The institute could support the development of appropriate testing methods, ensure the necessary access to models and prepare results for decisions by the German government and the competent authorities. Qualified independent testing organisations could carry out the technical evaluations and, in particular, assess whether the safety evidence for the use of AI in technical systems is reliable. Government responsibility and oversight would remain in place.

International cooperation with binding rules

From the perspective of the TÜV organisations, the national safety infrastructure must be embedded in a European and international framework. At EU level, they are calling for uniform safety requirements and the expansion of independent testing capacities. Together with partner countries such as the United Kingdom, Canada and Japan, Germany and the EU should coordinate testing procedures and negotiate jointly with developers on access to models and safety-relevant information.
The long-term objective is a binding international agreement providing for testing during development and before deployment, as well as clear limits for risks that cannot be controlled. If agreed risk thresholds are exceeded or compliance with key safety requirements has not been sufficiently demonstrated, development or deployment should be suspended. An international oversight structure should monitor compliance. A global reporting system for serious AI safety incidents should facilitate the exchange of information on emerging threats.

The joint white paper "Frontier AI Safety: Establishing Effective Governance Structures, Building an International Testing Framework" is available here.