TÜV Groups Propose Independent Testing for Frontier AI

In a joint announcement, TÜV-Verband and TÜV AI.Lab proposed independent safety testing for frontier AI models and autonomous AI agents at the German government's Digital Summit on November 2. Testing would begin during model development and continue throughout each model's life cycle.

Evaluators would examine model capabilities, safeguards and behavior under realistic operating conditions. They would need access to models, training processes and safety information, while specialists in machinery safety would assess AI systems that control physical equipment.

The groups propose expanding Germany's AISI into an independent institution with a clear mandate and dedicated resources. AISI would develop testing methods and provide findings to government authorities, while qualified inspection organizations would conduct technical evaluations.

The proposal also calls for common European requirements and coordinated testing with countries including the United Kingdom, Canada and Japan. Under a planned international framework, development or deployment would be suspended when models exceed agreed risk thresholds or fail to meet safety requirements.