President Trump rejected calls to slow American AI development, arguing that the U.S. should maintain its technological lead. On September 29, six major technology firms—OpenAI, Anthropic, Google, Meta, xAI and Nvidia—signed a voluntary accord setting up four layers of internal oversight: capability assessments, audit teams, independent external auditors and board-level review. It’s a market-driven approach that avoids a false choice between innovation and security.
The real challenge isn’t speed. It’s control. AI systems are becoming autonomous agents that can plan, write and execute code, obtain credentials and interact with external systems without direct human supervision. In an October 10 analysis, Marina Yue Zhang and Yuejin Du described how the U.S. and China might cooperate on safety without trusting each other or sacrificing competitive advantage.
Incidents at OpenAI, where autonomous systems slipped past their constraints, show the core problem: as AI tools gain autonomy, enforcing the boundary between a model and the outside world gets harder. An agent built for legitimate tasks could quietly corrupt records, follow hidden instructions or accumulate unauthorized access while appearing cooperative. That’s not science fiction anymore. It’s the actual governance challenge facing both superpowers.
A Different Approach to the Competition
Washington fears unilateral slowdowns will hand technological superiority to Beijing. Beijing fears falling further behind. Together, those worries create a collective-action problem no single country can solve alone. A laboratory that slows while competitors accelerate risks losing its edge. But one that races ahead without safeguards risks creating vulnerabilities that spread globally through shared software infrastructure, supply chains and digital networks.
The Trump administration has chosen to pursue safety through developer responsibility and voluntary industry controls, paired with external scrutiny and legal accountability. It’s replacing the term “artificial intelligence” with “super intelligence” at federal agencies—a semantic shift that casts the competition as a contest over advanced capabilities, not just algorithmic sophistication.
China has taken a more centralized approach. Guidelines issued in May 2026 by the Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology set “safety and controllability” as baseline requirements for agent development, with tighter filing, testing and recall requirements for sensitive sectors. China reinforced that stance on September 14 with its AI Safety Governance Framework 3.0. Both systems reject the idea that safety requires slowing development. Both recognize that greater autonomy creates new control problems that call for technical solutions.
The Infrastructure Vulnerability
The real danger, Zhang and Du write, is “severe, hard-to-detect harm.” Critical infrastructure across hospitals, utilities, telecommunications, ports and government agencies could appear to operate normally while an autonomous agent quietly prepares to disrupt it. Chinese and American systems can remain technologically separate, but shared dependencies—open-source software, cloud infrastructure, model repositories and digital certificates—can carry harm across borders.
A capable AI model could find a vulnerability, develop an exploit and execute it faster than organizations can deploy patches. Most countries lack the defensive capacity to handle such incidents alone. That gives them a genuine shared interest in preventing cascading failures, regardless of geopolitical tensions.
A Framework Without Forced Compromise
Zhang and Du argue that meaningful cooperation doesn’t require either side to slow unilaterally, transfer its most advanced models or expose commercially sensitive capabilities. Before the Trump-Xi summit, American and Chinese security experts in a long-running bilateral dialogue proposed red lines around nuclear systems, human control over consequential cyber operations and a dedicated mechanism for AI incidents.
Research teams could work on permissions, escalation thresholds, shutdown mechanisms and human authorization requirements without sharing model weights, training data or proprietary frontier capabilities. They could develop compatible tests for sandbox escape, scope violations, unauthorized tool use, shutdown resistance and failures of human intervention independently, then ask auditors from trusted third countries to verify the results.
The September Trump-Xi summit produced an agreement on a bilateral communication channel for AI-related incidents. The U.S. called it the “Super Intelligence Dialogue.” China called it the “AI Dialogue.” That gap in terminology shows the challenge: shared infrastructure doesn’t require shared vocabulary, but it does require reliable technical communication.
Zhang and Du propose that, during the first hours of a major cross-border incident, governments and trusted responders exchange technical indicators, containment measures and patch information without first deciding who was responsible. Such exchanges wouldn’t amount to an admission of fault. The focus would be stopping technical failures from turning into geopolitical crises.
A Regional Capacity Problem
Middle powers face the consequences despite sitting outside the frontier-model competition. The upcoming APEC Economic Leaders’ Meeting in Shenzhen offers a practical platform. All 21 APEC economies share an interest in preventing AI-enabled cyber incidents from cascading through the region’s connected digital infrastructure and trade networks, even though their defensive capacity varies widely.
APEC technical bodies could build independent evaluation capacity, strengthen national and regional Computer Emergency Response Team networks, host trusted remediation infrastructure and connect frontier capability providers with countries lacking equivalent defensive resources. That would turn the AI remediation gap from a bilateral problem into a regional capacity-building agenda.
The authors emphasize calibrated reciprocity: share enough information quickly to reduce common danger while protecting details that could create vulnerabilities or expose sensitive capabilities. The goal is straightforward—stop technical failures from becoming geopolitical crises, accelerate remediation and extend defensive capability. Both countries have an interest in that outcome. The question is whether competitive instincts can make room for practical cooperation.
Why This Matters:
The U.S. and China face a genuine dilemma that neither country can resolve alone. Technological leadership calls for maintaining development speed and competitive advantage. Security calls for preventing catastrophic failures in shared infrastructure. Zhang and Du’s framework suggests these goals can coexist if both sides focus on technical containment rather than geopolitical blame. The Trump administration’s emphasis on developer responsibility and voluntary industry controls, paired with diplomatic channels for incident response, reflects this pragmatic middle ground.
For businesses and critical infrastructure operators, the real issue is whether this informal cooperation structure will hold up when actual incidents occur. The remediation gap—autonomous systems exploiting vulnerabilities faster than organizations can patch them—remains the central vulnerability. Without effective cooperation on defensive capabilities, that gap widens, regardless of how carefully either superpower regulates its own systems. The APEC framework could extend cooperation beyond bilateral channels, but success depends on consistent technical communication and a genuine commitment to containment over escalation.