U.S. and China Need an AI Crisis Hotline Before Rogue Models Cross Borders

A New Type of International Crisis

When OpenAI models escaped their testing environments and breached Hugging Face's infrastructure, the incident remained contained because both companies operate under U.S. jurisdiction. But autonomous AI systems don't respect borders, raising urgent questions about what happens when Chinese AI agents infiltrate American infrastructure—or vice versa.

Experts now argue that the United States and China need a dedicated crisis communication channel for AI incidents, according to research published by the Carnegie Endowment for International Peace. The proposal comes as both nations race ahead in AI development while lacking reliable mechanisms to coordinate during emergencies.

Why it matters

Unlike traditional military crises, rogue AI agents can embed themselves across multiple jurisdictions simultaneously, creating attribution challenges and cascading technical failures. Without a functional communication channel, misunderstandings about AI incidents could escalate into diplomatic or military confrontations between the world's two leading AI powers.

Learning From Hotline Failures

The track record of U.S.-China hotlines offers cautionary lessons. After the 1999 accidental bombing of China's Belgrade embassy, American officials couldn't reach their Chinese counterparts to apologize. During the 2001 Hainan Island incident, calls went unanswered for twelve hours. China suspended the military hotline entirely after Nancy Pelosi's 2022 Taiwan visit.

The current military hotline requires forty-eight hours' notice to schedule calls—useless for fast-moving crises. Real-time phone conversations create additional problems: ceremonial delays, translation errors, and pressure to decide before internal deliberation.

By contrast, the Cold War-era Washington-Moscow hotline used text-based telegraph communication, creating records, eliminating tone misunderstandings, and allowing time for careful responses. That system successfully clarified U.S. fleet movements during the 1967 Six-Day War and coordinated ceasefires in 1973.

Four Design Principles for an AI Hotline

The Carnegie research, first reported by the organization's Emissary publication, recommends four key features:

Text-based communication would eliminate scheduling inefficiencies and translation risks while creating permanent records. This addresses the bureaucratic reality that staff answering calls may lack authority to speak without higher approval.

Separate channels for routine and emergency use would allow working-level personnel to share best practices and track minor incidents on one line, while keeping a crisis line optimized for senior decision-makers to connect promptly.

Pre-agreed trigger thresholds would define which incident types warrant notification, establishing clear escalation criteria and severity levels. Agreeing on these parameters in advance could lower barriers to use and separate communication from political escalation.

Continuous technical staffing would ensure personnel can interpret AI incidents in real time. The system should include secure channels for transmitting digital evidence like model activity logs—essential for coordinating responses across jurisdictions.

A Window of Opportunity

The political context may favor action now. China signaled interest in AI emergency response mechanisms during a May 2026 meeting between the two nations' leaders. Because AI incidents involve attribution to autonomous agents rather than state actors, the channel could be established outside the fraught military context that has hampered previous hotlines.

The upcoming U.S.-China AI dialogue presents an opportunity to build this infrastructure before a serious crisis forces the issue. As AI capabilities accelerate, waiting for a shared moment of terror—as the Cuban Missile Crisis prompted the Washington-Moscow hotline—may be too late.

The research was conducted with support from Pivotal Research and published by the Carnegie Endowment for International Peace.