German Foreign Minister Johann Wadephul called for common safety requirements for the most capable AI models at the UN General Assembly on September 26, 2026. Providers should have to test models before release and report serious safety failures. He said the requirements should be enforced through national or regional oversight.
Wadephul framed the proposal around who makes decisions about a technology with worldwide consequences and who is accountable. He warned about risks to infrastructure and public trust, while also pointing to possible benefits for research, health and education.
The minister called for stronger public capacity to assess powerful AI and for every region to help shape common rules. His argument rejects reliance on voluntary pledges alone. The speech calls for more transparency and shared standards, while retaining the industry's responsibility for developing its systems.
The appeal is not an international agreement. It specifies neither a testing procedure nor an implementation timetable. We record the officially published speech as a policy Risk Signal: a contribution to the debate about verifiable control, without claiming a further AI incident or a protection already in force.