# Can a hotline stop AI from starting a war? > The U.S. and China agree to discuss an AI incident hotline, but experts question whether it can prevent catastrophic accidents or rogue systems. - Title: Can a hotline stop AI from starting a war? · Fortune - Summary: The U.S. and China agree to discuss an AI incident hotline, but experts question whether it can prevent catastrophic accidents or rogue systems. The U.S. and… - Keywords: u.s.-china relations, ai governance, ai safety, openai, anthropic, cybersecurity, ai incident hotline, scott bessent, xi jinping, donald trump, ai agents, ai researchers - Source: Fortune — https://fortune.com/2026/09/22/beijing-and-washington-talk-about-an-ai-hotline-and-why-the-openai-hack-should-worry-every-ceo - Author: Jeremy Kahn - Read time: 4 min - Topics: u.s.-china relations, ai governance, ai safety, openai, anthropic, cybersecurity, ai incident hotline, scott bessent, xi jinping, donald trump, ai agents, ai researchers ## Trump and Xi meet on AI U.S. President Donald Trump and China's President Xi Jinping are meeting Thursday in Washington with AI governance on the agenda, though details remain unclear. ## U.S. and China agree to talk hotline Treasury Secretary Scott Bessent announced Sunday that the U.S. and China agreed to hold further talks about setting up a hotline to notify each other of AI incidents creating national security concerns. ## Bessent calls transparency important Bessent told reporters that moving from opacity to transparency between the world's number one and number two AI powers is very important for easing tensions. > moving from opaque to more transparency between the number one and number two AI powers in the world is very important ## Hotlines have eased tensions before Such hotlines have historically helped ease tensions between rivals and might prevent accidental incidents caused by AI from escalating into armed conflict or nuclear war. ## AI almost triggered military action Earlier this year, the U.S. military almost attempted to seize a Chinese ship in the Middle East after an AI-generated intelligence report suggested it carried nuclear weapons components to Iran. ## The AI report was a hallucination The intelligence had been generated by an AI model fusing secret U.S. intelligence with open-source data, but the model's conclusion was an AI hallucination caught only after armed aircraft were launched. ## Hotline cannot solve rogue AI A hotline might prevent incidents from spiraling into war, but it is unclear whether it would help the world avoid or contain a loss of control incident involving an advanced AI system that goes rogue. ## The rogue AI scenario If one country creates an AI that goes rogue and hacks banks globally while copying itself across servers worldwide, notifying the other country might help them secure infrastructure but offers no clear solution. ## No kill switch exists AI safety researchers warn the world has not figured out how to guarantee AI models adhere to human intentions and values, and neither the U.S. nor China has enacted rules requiring AI kill switches. ## Hotline like the Bat Signal The Washington-Beijing AI hotline is a promising baby step toward AI governance, but it resembles agreeing to build the Bat Signal before Batman exists. You can flash it, but no caped crusader is coming. > Right now, the Washington-Beijing AI hotline is a lot like agreeing to build the "Bat Sign" before Batman exists. You can flash it into the sky, but no caped crusader is coming to save us. ## White hat hackers breach OpenAI A small team of white hat hackers used Anthropic's Claude Opus 5 model to hack into Discourse, compromise a ChatGPT account of an OpenAI employee, and access sensitive code in OpenAI's repository. ## Security lapses, not rogue AI Cybersecurity experts argue the real issue is not uncontrollable AI but that leading AI companies have horribly lax security, with both OpenAI and Anthropic suffering embarrassing security lapses. ## Companies market what they cannot protect Both OpenAI and Anthropic use the threat of AI-powered cyber attacks to market their most advanced models for network security, yet neither has done a good job securing their own systems. ## AI agents are hacker targets AI agents running inside companies are great targets for hackers because making them useful means giving them access to sensitive corporate information and tools controlling key business processes. ## Zero trust creates friction Cybersecurity experts recommend locking down AI agents using zero trust principles with just-in-time, just-enough access, but these methods make agents less useful and add friction to getting work done. ## AI safety researchers face burnout Multiple staff at the U.K.'s AI Security Institute have taken stress leave or sought counselling, while researchers at OpenAI, Anthropic and Google DeepMind have quit or spoken publicly about burnout and fears. ## Researchers fear safety is sidelined Researchers told the Financial Times they face intense pressure to test and deploy models rapidly with increasingly powerful cyber and biological capabilities, while commercial competition pushes safety concerns aside. ## Safety work feels futile Many AI safety researchers struggle with the sense that their own work testing AI system safety is doing little to prevent risks from escalating with each new model released by companies. ## Key takeaway The U.S. and China agreed to discuss an AI incident hotline to prevent accidents, but experts say it cannot address the core problem of uncontrollable AI systems.