US-China AI Rivalry Strains Global Safety Cooperation

Tensions reached a new peak after U.S. officials alleged that Chinese startup Moonshot used outputs from Anthropic’s Fable 5 model to train its own Kimi K3 system. Treasury Secretary Scott Bessent has since raised the prospect of sanctions against Chinese developers, while the Commerce Department probes potential illegal access to restricted American semiconductors. These friction points arrive just as experts warn that advanced models are nearing autonomous capabilities that could be exploited for cyberattacks or critical infrastructure disruption.

Despite the hostility, cooperation remains a strategic necessity. Researchers emphasize that as AI models gain recursive self-improvement features, the absence of shared international standards leaves the global community vulnerable. Yoshua Bengio has cautioned that once highly capable open-weight models are widely distributed, mitigating misuse becomes nearly impossible without rigorous, government-backed evaluation systems. Currently, China mandates state-led content reviews, while the U.S. relies largely on voluntary evaluations through the Center for AI Standards and Innovation.

Internal debate in Washington complicates the path forward. White House adviser David Sacks has criticized calls for aggressive restrictions, suggesting that some domestic firms are seeking regulatory barriers to stifle lower-cost Chinese competition. With Chinese models reportedly accounting for 60 percent of token usage on the OpenRouter platform, the economic stakes are rising alongside security concerns. Whether the planned September dialogue survives these political pressures will determine if the two powers can establish common red lines or if they will drift toward entirely separate, opaque regulatory frameworks.