September 23, 2026

SHARE |

A Hotline for the AI Age: The U.S.-PRC Imperative to Cooperate on AI Safety and Prevent Terrorist Misuse

P Photo/Andy Wong

Last weekend, after U.S. Treasury Secretary Scott Bessent and PRC Vice Premier He Lifeng met in New York City, Bessent announced that the two parties would establish a formal dialogue on AI safety, including an incident-sharing mechanism for relevant AI safety incidents. According to Bessent, the two sides will start discussing “protocols … so both sides can agree on what the leading AI dangers are, whether it's uncontrollable agents, whether it’s non-state actors, and cyber non-state actors.” A new Insight brief by the Soufan Center, Terrorist Misuse of AI: A Framework for U.S.-PRC Cooperation, advocates prioritizing terrorist misuse of AI in this bilateral dialogue. For such AI safety dialogue to succeed amid strategic competition between Beijing and Washington, terrorist misuse is a logical starting point, as the potential harm is a high priority, the issue commands broad consensus, and the domain is relatively depoliticized. This specifically applies to misuse by groups considered national security threats by both capitals, including al-Qaeda, Islamic State, and their supporters. This is appropriate as both Washington and Beijing continue to lead AI development, whether it is the frontier proprietary models developed in the United States or the rapidly improving open-source and open-weight models developed in the PRC.

A framework should be narrowly scoped to avoid running afoul of other U.S. priorities and values. Concretely, this means safeguarding the cooperation framework from collapsing at the next instance of disagreement over semiconductor export controls or other strategic-competition-related challenges, and from becoming a vehicle for domestic political considerations in the PRC, such as the surveillance and prosecution of Uyghurs in Xinjiang. The Insight proposes five pillars to operationalize cooperation, including (i) a working group on mitigating terrorist misuse of AI, (ii) Chemical, Biological, Radiological, and Nuclear (CBRN) cooperation, beginning with nucleic acid synthesis screening, (iii) Track II dialogue on risk reduction of terrorist AI misuse, (iv) parallel national threat taxonomies and the establishment of a crosswalk, and (v) the establishment of an incident information sharing mechanism. For such a cooperation framework to succeed, the parties would need to agree on guardrails at the outset, necessarily excluding any requirement to disclose model capabilities, weights, training data, or evaluation results.

The time is more than ripe for cooperation on this issue. Terrorist misuse of AI — specifically that by Islamic State, al-Qaeda, and their supporters — has matured beyond experimentation. Initially primarily used for propaganda operations and meddling in the information environment, both generative and agentic AI capabilities are increasingly part of operational planning and have been noted in thwarted plots globally. This aligns with the atomization of the terrorism threat, encompassing a surge in individuals without formal ties to terrorist groups responsible for terrorist plots and political violence, eclipsing the traditional form of external operations by hierarchical groups like Islamic State Central and al-Qaeda Central. Often, such lone actors resort to simple attack methods, like do-it-yourself (DIY) guides, and now, personal advice is given from chatbots on anything from reaching a terrorist group in a conflict zone to the production of ricin. At the same time, generative AI adoption has been observed among terrorist groups, including Islamic State West Africa Province, for operational and organizational needs. The rapid advancement in capabilities across both proprietary and open-weight models suggests adoption may drive real uplift by facilitating propaganda, operations, and operational security practices, as well as supplying novel information and capabilities or removing previously faced bottlenecks.

The imperative to prevent terrorist misuse of AI is urgent and demands a considered, pragmatic solution. The U.S. and PRC, as frontrunners in frontier model development, are well placed to tackle this issue set. Nonetheless, in the context of great power competition, challenges abound, and AI development and deployment have become a frontier of strategic competition. In 2025, the Trump White House published “America’s AI Action Plan,” which stated: “The United States is in a race to achieve global dominance in artificial intelligence (AI). Whoever has the largest AI ecosystem will set global AI standards and reap broad economic and military benefits.” The Chinese Communist Party (CCP) has emphasized the need for China to become “the world’s primary AI innovation center,” to “seize the commanding heights of science and technological development,” and to “resolutely win the battle over key and core technologies.” Additionally, Chinese perspectives on safety and governance have often prioritized regime stability. They have also downplayed the distinct misuse opportunities presented by open-weight and open-source AI models, since those stand in conflict with the PRC’s open-source AI strategy. The PRC also continues to define AI safety expansively, including issues such as U.S. export controls as a “supply chain safety risk.” Safety thus relates not only to model behavior and access, but also to regime security.

Regardless of these substantial challenges, there is a shared interest in preventing non-state actor misuse of these models, including terrorists and violent extremists in both countries. The PRC’s AI Safety Governance Framework 3.0, released earlier this month, indicates continued concern that AI lowers the barrier to knowledge of nuclear, biological, chemical, and missile weapons, which could be “maliciously exploited by criminals, extremist groups, or terrorists.” This lower-probability, higher-impact risk has dominated the AI safety risk domain, and both capitals have independently highlighted the potential uplift non-state actors may receive from AI related to this specific misuse risk.

Catastrophic proliferation to non-state actors, including terrorist groups, is a symmetric downside for both countries: neither Washington nor Beijing gains from those capabilities becoming widely accessible, and neither can contain the consequences within its own borders. While other AI risk issues appear at times too contentious to spark sustained dialogue that can survive an incident in bilateral relations, a narrowly scoped and pragmatic framework on terrorist misuse can stand the test of time.