In view of the catastrophic risks associated with increasingly powerful AI models, US President Donald Trump and Chinese President Xi Jinping may discuss the establishment of a bilateral communication channel for AI incidents at this week’s Asia-Pacific Economic Cooperation summit in Shenzhen, China. The Nuclear Risk Reduction Center (NRRC), established in 1987 based on a US–Soviet agreement, is often brought up as a precedent and model for a communication channel between adversaries. The current US and Russian NRRCs represent a 24/7 open and secure communication channel to reliably, rapidly, and securely transmit notifications and other messages related to the implementation of arms control treaties and agreements.
While this model has merits, it also has limitations. An examination of the merits and limitations is an opportunity to unmask hidden potential for a more meaningful debate on AI and international security. Such a debate would recognize that an NRRC model is pointless if not accompanied by state regulation of AI companies and corporate accountability structures, given the fact that frontier (or the most advanced) AI research, unlike nuclear weapons in the government’s possession, resides in the private sector. It would also emphasize cooperation and capacity-building on AI among multiple industry and state stakeholders. Finally, it would attend to the myriad ways AI is implicated in international security, from nuclear proliferation and military conflict escalation to the language of an “AI race” and the consequences of massive industry spending and government allocation of resources.
Background: Trump and Xi’s September summit
The September 23–25, 2026, summit between Trump and Xi was much anticipated. On the agenda were trade, the US–Israel war on Iran, and Taiwan, but also AI. The AI agenda item could not have been timelier: There had been a relentless stream of news about cybersecurity breaches by frontier AI models since July (including some breaches that only came to light months afterwards), documentation of Anthropic models being misused several times in less than a year, whistleblower accounts of insufficient safety research efforts by AI developers, and an Anthropic call for “pacing the frontier”—that is, slowing AI progress—that was affirmed by other AI company CEOs.
In the same week, diplomats, dignitaries, and tech experts at the 81st session of the United Nations General Assembly called for rogue AI to be reined in. For its part, the Trump administration admonished AI developers both for their “doomerism”—the belief that AI has a significant chance of causing human extinction—and “globalist governance” proposals, citing a need to outpace China as a reason to charge ahead in pursuit of advanced AI while being subject only to the lightest regulation.
Trump and Xi established the US–China Super Intelligence (SI) Dialogue to exchange views on risks and benefits related to AI, and agreed to establish a bilateral communication channel on AI incidents. All eyes will be on whether this week’s exchange will achieve progress on a mechanism for making that channel a reality. The United States and China could use the NRRC as a template as long as they take both its merits and limitations into account.
Transparency in high-risk settings could prevent catastrophe
The NRRC model’s merits include recognition of the need for transparency in the actions of adversarial states that could trigger a nuclear war, and providing a means to communicate, clarify, and de-escalate.
AI models, which also pose catastrophic risks, may benefit from similar communication. For example, US AI models are gaining exponentially in cyber capabilities and can identify vulnerabilities and the means to exploit them at a higher speed and efficiency than previously existed. A catastrophic risk would be if AI agents, in part due to lax lab safety measures, hack into an adversary’s critical infrastructure systems. In today’s highly networked societies, such an attack could wreak economic and social chaos. But there is also the danger of an attack on already entangled military or nuclear command, control, and communications systems, which might lead to nuclear escalation. The current 2022 US Nuclear Posture Review suggests that the United States might use nuclear weapons not only in a potential nuclear attack by an adversary, but also in non-nuclear “high consequence, strategic-level attacks” by an adversary, which may involve cyber and other advanced conventional capabilities.
The NRRC’s formatted short-form communication could have the benefit of forcing states to start thinking about technical, procedural, and organizational issues—thinking that might culminate in broader domestic, bilateral, and international AI governance frameworks.
The limits of a state-centric perspective
A limitation of the NRRC model if applied to AI is that it was designed to provide state-to-state notifications on state actions related to arms control agreements and treaties. So, the relevant information was generated, understood, and communicated by the entity responsible for those actions. While governments exercise jurisdiction over companies in their territory, US regulation of AI models—and thus its ability to spot problems—has been light, with companies merely encouraged to release models to the government on a voluntary basis and with little in place in terms of genuinely independent government testing, evaluation, and monitoring mechanisms. Some cyberhacking incidents came to light recently only when companies self-reported.
That points to another danger of the state-centric NRRC framework: Corporate accountability is deprioritized, obscuring safety and security failures that are squarely the responsibility of the developer. This lack of accountability allows developers to perpetuate their preferred narratives that the United States must win the “AI race” with China and that rogue machines exceed our human capacity to understand and control them; in this way, developers can attribute their own failures to the advanced state of their creations rather than having to admit to any missteps in their practices and accept any responsibility for the consequences. While calls for improved and rigorous liability mechanisms for AI failures are bipartisan and increasingly loud, none have been developed as yet.
Lessons from the evolution of the NRRC model
Proposals to create an NRRC-like mechanism for AI should avoid the narrow focus of the original and incorporate two significant historical developments. The first is that the US NRRC expanded and evolved over time, and the international efforts that followed implemented this expanded function. The current scope of what is now called the US National and Nuclear Risk Reduction Center (NNRRC) is “nuclear arms control, conventional arms transparency, [chemical weapon] CW destruction, ballistic missile and space launch vehicle awareness, and international cyber incidents.” The set of communication partners has expanded beyond Russia to include other nations and organizations including the Organization for Security and Co-operation in Europe and the Organisation for the Prohibition of Chemical Weapons. (See the State Department’s History of the National and Nuclear Risk Reduction Center.)
The second historical development that could provide inspiration for an AI template is the United Nations effort that emulates the cyber arm of the NNRRC, the Global Intergovernmental Points of Contact Directory on the Use of Information and Communications Technologies in the Context of International Security. Both of these efforts are accompanied by information sharing, cooperation, and, in the UN case, capacity-building activities.
The precedents of expanded concept and connected activities have rich potential for advancing AI risk reduction. In contrast, if we fixate on the relationship between the NRRC and nuclear weapons, we risk perpetuating an arms race frame that would prioritize accelerated AI development over security. We would also miss out on lessons from non-nuclear domains that might be more relevant and constructive, as the state-centric nuclear/AI analogy that has animated many pens meets its limits.
AI security implications are broader than those implied by nuclear analogies
Cybersecurity concerns gesture to the need to think expansively about security. If the NRRC is to be viewed as a template for AI security, we should be asking what security aspects might be obscured by this model.
One aspect does link back to nuclear energy and weapons. In AI developers’ rush to achieve artificial general intelligence (or maybe primarily to maximize profits), they have contributed to a nuclear energy renaissance. Nuclear energy is touted by the companies as necessary to power data centers’ servers and cooling equipment while meeting self-declared carbon emissions goals. The drive to revive and build out nuclear power generation projects comes at an opportune time for that industry, when the US government is relaxing its regulatory environment to accommodate these initiatives. Thus, the AI rush may actually increase nuclear risks by spurring a deregulated nuclear energy industry, which could have effects beyond US borders. The Trump administration’s proposed nuclear energy deal with Saudi Arabia, for example, features weakened nonproliferation measures. (And none of this speaks to the actual viability of new nuclear power projects.)
There are (at least) two other securitization aspects we should consider. AI may underlie autonomous weapon systems, defined in the US Department of Defense Directive 3000.09 2023 update as systems that “once activated, can select and engage targets without further intervention by an operator.” There is widespread conviction that the greatest risk of nuclear war comes from escalation from conventional war. AI-driven autonomous weapon systems increase the tempo of warfighting and exacerbate the fog of war, leading to mistakes and raising the risk of nuclear escalation. Given the increased tempo and uncertainty, this is likely to be the case even with humans not completely eliminated from observation, orientation, decisionmaking, and action loops. And yet, military AI regulation efforts have been slow to mature: The crowning achievement after years of US–China Track 1.5 and 2 dialogues is a glossary of AI terms, while systems of high lethality are already in use and have caused great loss of life by Israel in Gaza and by the United States in Iran.
Finally, the language of an “AI race” and related practical measures elevate AI to a strategic asset, so much so that data centers have come to be seen as attractive military targets by their mere association with AI. This can be seen in Iran’s attacks on Amazon data centers in the United Arab Emirates and Ukrainian attacks on Russian data centers.
The narrative of AI as a strategic asset is reinforced by export controls targeting China on advanced chips and chipmaking materials and technologies, and by excessive US government investment in the AI sector. This investment may be a show of faith, but it could also be a preemptive bailout for a market built mainly on debt and equity—a market speculated by some to be a bubble about to burst—with little to show in terms of actual returns other than the public’s ire and an absence of federal regulation.
These broader security implications of AI are not neatly captured by proposals to use the NRRC as a template, and may be obscured if these efforts are too closely focused on the nuclear weapons aspects of the NRRC. Keeping in mind the history of changes to the US NRRC and other efforts, Presidents Trump and Xi have an opportunity to begin pushing risk reduction efforts in a more constructive direction, while hopefully spreading genuine AI benefits internationally.
