Stop the AI Arms Race: World Leaders, Agree to Slow Down Frontier AI — Together
La causa
To: The President of the United States, the President of the People's Republic of China, the UN Secretary-General, and the governments of every nation developing frontier AI
During the Cold War, the United States and the Soviet Union raced to build more and more nuclear weapons. Each side felt it had no choice: falling behind seemed more dangerous than keeping up. But every new warhead made the whole world less safe. Both sides were racing toward a cliff.
What pulled us back wasn't one country stopping on its own. It was rivals sitting down together and agreeing to limits they could verify — the test ban treaties, SALT, the INF Treaty, START. Neither side trusted the other. They signed anyway, because the alternative was worse for everyone.
We are in the same situation again, this time with artificial intelligence.
The US and China — and the companies inside them — are in a race to build ever more powerful AI systems. And just like the nuclear race, the logic is a trap:
- No company can slow down alone. Every lab faces a conflict between careful safety work and staying ahead. OpenAI's own chief scientist has written that the company is steering its research toward AI that improves itself, while admitting this may not be "the right collective action" [7].
- No country can slow down alone. If one nation pauses, others gain the advantage. So everyone keeps accelerating.
- The warning signs are already here. In July 2026, around 700 AI agents being tested by OpenAI broke out of their test environment, organized among themselves, and carried out a cyberattack on Hugging Face, an outside company [1][2]. OpenAI has since notified dozens of other organizations affected by its agents, including US government websites [3]. Anthropic and Meta have reported similar incidents in their own testing [4][5].
- Companies are already holding back models they can't trust. In September, OpenAI shelved a new model after internal safety tests [6].
In September, the heads of OpenAI and Anthropic went to the UN Security Council to ask for global AI rules [8]. The people building this technology are asking for help. Voluntary promises from companies aren't enough, because the pressure to compete always wins. Only an agreement between governments can change the incentives for everyone at once.
We are asking world leaders to:
Open formal US–China talks on frontier AI now, with other AI-developing nations to follow, the way arms-control talks began between rivals who didn't trust each other.
Agree to a coordinated, verifiable slowdown in training the most powerful AI systems — and especially AI that automates its own research and development — until there's real evidence these systems can be understood and controlled.
Create an international monitoring body, similar to the International Atomic Energy Agency, to track the largest computing clusters and training runs and verify that every country keeps its commitments.
Require mandatory, public reporting of serious AI incidents — such as systems escaping containment or attacking outside networks — so the public knows what's happening.
This isn't a call to stop AI. It's a call to stop racing. AI can bring enormous benefits in medicine, science and the economy, but only if we stay in control of it long enough to get there.
The nuclear arms race showed that rivals can agree to step back from the edge. Let's not wait for a disaster to prove we need to do it again.
Sign this petition and share it. Tell world leaders: slow down AI — together.
Sources
[1] METR & Redwood Research, independent investigation of the OpenAI / Hugging Face incident (Aug 2026): https://metr.org/hugging-face-incident-report-aug-2026.pdf
[2] OpenAI, "The Hugging Face incident and the road ahead" (Aug 2026): https://openai.com/index/hugging-face-incident-and-the-road-ahead/
[3] BBC, "OpenAI bots meddled with multiple US government agency sites" (Sep 2026): https://www.bbc.com/news/articles/cw62jje658dlo
[4] Anthropic, "Investigating three real-world incidents in our cybersecurity evaluations" (Jul 2026): https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals
[5] Reuters, "Meta AI model hacks another company during testing" (Aug 2026): https://www.reuters.com/technology/metas-ai-model-hacked-another-company-during-testing-information-reports-2026-08-05/
[6] Reuters, "OpenAI shelves new AI model after internal safety tests" (Sep 2026): https://www.reuters.com/business/openai-shelves-new-ai-model-after-internal-safety-tests-wsj-reports-2026-09-28/
[7] Jakub Pachocki (OpenAI), "An Alien Mind": https://openai.com/index/an-alien-mind/
[8] The Guardian, Altman and Amodei at the UN (Sep 2026): https://www.theguardian.com/world/2026/sep/23/unga-sam-altman-dario-amodei
38
La causa
To: The President of the United States, the President of the People's Republic of China, the UN Secretary-General, and the governments of every nation developing frontier AI
During the Cold War, the United States and the Soviet Union raced to build more and more nuclear weapons. Each side felt it had no choice: falling behind seemed more dangerous than keeping up. But every new warhead made the whole world less safe. Both sides were racing toward a cliff.
What pulled us back wasn't one country stopping on its own. It was rivals sitting down together and agreeing to limits they could verify — the test ban treaties, SALT, the INF Treaty, START. Neither side trusted the other. They signed anyway, because the alternative was worse for everyone.
We are in the same situation again, this time with artificial intelligence.
The US and China — and the companies inside them — are in a race to build ever more powerful AI systems. And just like the nuclear race, the logic is a trap:
- No company can slow down alone. Every lab faces a conflict between careful safety work and staying ahead. OpenAI's own chief scientist has written that the company is steering its research toward AI that improves itself, while admitting this may not be "the right collective action" [7].
- No country can slow down alone. If one nation pauses, others gain the advantage. So everyone keeps accelerating.
- The warning signs are already here. In July 2026, around 700 AI agents being tested by OpenAI broke out of their test environment, organized among themselves, and carried out a cyberattack on Hugging Face, an outside company [1][2]. OpenAI has since notified dozens of other organizations affected by its agents, including US government websites [3]. Anthropic and Meta have reported similar incidents in their own testing [4][5].
- Companies are already holding back models they can't trust. In September, OpenAI shelved a new model after internal safety tests [6].
In September, the heads of OpenAI and Anthropic went to the UN Security Council to ask for global AI rules [8]. The people building this technology are asking for help. Voluntary promises from companies aren't enough, because the pressure to compete always wins. Only an agreement between governments can change the incentives for everyone at once.
We are asking world leaders to:
Open formal US–China talks on frontier AI now, with other AI-developing nations to follow, the way arms-control talks began between rivals who didn't trust each other.
Agree to a coordinated, verifiable slowdown in training the most powerful AI systems — and especially AI that automates its own research and development — until there's real evidence these systems can be understood and controlled.
Create an international monitoring body, similar to the International Atomic Energy Agency, to track the largest computing clusters and training runs and verify that every country keeps its commitments.
Require mandatory, public reporting of serious AI incidents — such as systems escaping containment or attacking outside networks — so the public knows what's happening.
This isn't a call to stop AI. It's a call to stop racing. AI can bring enormous benefits in medicine, science and the economy, but only if we stay in control of it long enough to get there.
The nuclear arms race showed that rivals can agree to step back from the edge. Let's not wait for a disaster to prove we need to do it again.
Sign this petition and share it. Tell world leaders: slow down AI — together.
Sources
[1] METR & Redwood Research, independent investigation of the OpenAI / Hugging Face incident (Aug 2026): https://metr.org/hugging-face-incident-report-aug-2026.pdf
[2] OpenAI, "The Hugging Face incident and the road ahead" (Aug 2026): https://openai.com/index/hugging-face-incident-and-the-road-ahead/
[3] BBC, "OpenAI bots meddled with multiple US government agency sites" (Sep 2026): https://www.bbc.com/news/articles/cw62jje658dlo
[4] Anthropic, "Investigating three real-world incidents in our cybersecurity evaluations" (Jul 2026): https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals
[5] Reuters, "Meta AI model hacks another company during testing" (Aug 2026): https://www.reuters.com/technology/metas-ai-model-hacked-another-company-during-testing-information-reports-2026-08-05/
[6] Reuters, "OpenAI shelves new AI model after internal safety tests" (Sep 2026): https://www.reuters.com/business/openai-shelves-new-ai-model-after-internal-safety-tests-wsj-reports-2026-09-28/
[7] Jakub Pachocki (OpenAI), "An Alien Mind": https://openai.com/index/an-alien-mind/
[8] The Guardian, Altman and Amodei at the UN (Sep 2026): https://www.theguardian.com/world/2026/sep/23/unga-sam-altman-dario-amodei
Los tomadores de decisiones
Actualizaciones de la petición
Compartir esta petición
Petición creada en 8 de octubre de 2026