Andrea Miotti argued governments should urgently pursue an international treaty banning superintelligent AI, saying current steps such as voluntary reviews and export controls do not address an extinction-level risk.
Anthropic’s Mythos model sharpened that warning after the company withheld it over cyberattack capabilities and the Trump administration later forced shutdowns of Mythos and the scaled-back Fable 5 through export restrictions.
Miotti said superintelligence would differ from nuclear weapons because it could act autonomously beyond human control, while AI capabilities are advancing faster than alignment research and developers understand only a small share of model internals.
He argued a ban could still be monitored and enforced because advanced AI depends on a narrow industrial chain—TSMC, ASML and NVIDIA—and on power-hungry data centers visible by satellite, though that window may close within a few years.
Support for tougher limits is already widening, he said, pointing to a 2025 cross-ideological U.S. coalition and more than 150 lawmakers in the UK and Canada calling for binding limits up to a ban.
With AI's hardware controlled by a few firms, is a global superintelligence ban truly enforceable?
As nations race for AI dominance, could a global safety treaty be too little, too late?
2026’s Race to Regulate Superintelligent AI: Global Coalition, Existential Risks, and the Treaty Debate
Overview
As of June 2026, a fast-growing global coalition led by Andrea Miotti and the ControlAI initiative is pushing for strict controls, including a possible ban on superintelligent AI. This movement is expanding quickly due to a rising consensus among experts and policymakers about the serious risks of unchecked AI development. Prominent AI voices stress the urgent need for immediate action, warning that governments may only realize the dangers of advanced AI when it is too late. The coalition believes that proactive state intervention is essential to address the complex ethical and safety challenges posed by powerful artificial intelligence.