BELGRADE—Much of the focus on international safety discussions around artificial intelligence (AI) has understandably centered on the bilateral US-China relationship. Technological competition between the United States and China indeed defines the parameters of what is possible in AI regulation. However, any workable global AI safety governance will need to include Europe as well, especially the European Union (EU) and its massive common market.
At the same time, digital commerce has frequently been an irritant in the transatlantic economic relationship in recent years. To avoid further friction between the world’s largest trade and investment partners—and to address the potentially catastrophic risks of a technology with massive financial promise—policymakers on both sides of the Atlantic should explore ways to coordinate more on AI safety.
Building on existing programs
Unfortunately, the Trump administration has so far shown little interest in engaging with European allies on AI safety. That could change. Recent cybersecurity incidents emanating from US frontier labs may open a track for narrower coordination on some cyber or biological risks while avoiding more contentious issues for the administration, such as content moderation or data privacy. The White House could lead an effort to increase transatlantic dialogue on AI, but progress is more likely to come from US and European legislatures, which could readily build on existing programs. For example, Washington and Brussels could create a joint AI safety working group within the existing Transatlantic Legislators’ Dialogue (TLD) between the US House of Representatives and the European Parliament. Given the acute risk from AI-enabled cyberattacks, this group should begin with a narrow focus on cybersecurity measures while aiming to expand its AI safety coordination over time.
Since 1999, the TLD has met twice a year, alternating between the United States and Europe. The dialogue has provided a format for inter-parliamentary exchange on trade and security issues. Recent joint statements have called for increased US-EU cooperation on AI and other emerging technologies. The European Parliament’s TLD activities are coordinated by a steering committee that includes the chairs of the US delegation, the Foreign Affairs Committee, and the International Trade Committee. The US delegation is an ad hoc bipartisan group led by a co-chair and a ranking member.
The TLD’s flexible structure makes it well suited for an AI safety-focused session next year, when the new Congress convenes. With AI safety a priority for the incoming Democratic majorities, an early dialogue with EU partners would help to ensure any potential new enforcement measures are interoperable. In public opinion polls, Americans tend to express more concern about AI risks than Europeans, but pluralities across Europe express ambivalent feelings toward the technology. Exit polling from the US midterm elections shows legislators face pressure from constituents to deliver meaningful AI safety measures, which will need to include risks that extend across borders.
Currently, US AI safety policy remains a patchwork of voluntary commitments, executive actions, and state legislation. In Congress, several draft bills advancing AI safety, such as the bipartisan FRONTIER Act, have already been introduced or are circulating informally. The October 2026 White House Accord on Super Intelligence also established voluntary commitments from leading AI companies. Meanwhile, the EU has adopted the AI Act as its legal framework for AI regulation, but recent amendments to the act this summer via the Digital Omnibus, which entered into force in July 2026, adjusted implementation timelines and compliance requirements, showing that the book is not yet closed. Important questions remain unresolved, like oversight of models released to a restricted group of users, such as Anthropic’s limited rollout earlier this year of Mythos through Project Glasswing.
Avoiding failure, achieving success
Of course, any new US-EU effort should be clear-eyed about potential points of failure. Previous transatlantic dialogues have faltered, for instance, when one side perceived the other as attempting to export its regulatory regime. Another common pitfall has been a lack of tangible results to show constituents, which has led to US-EU dialogues often being dismissed as mere talk shops. An agenda focused on the five short- to medium-term deliverables below could avoid such failures:
- Interoperable incident reporting: Both sides should adopt a common taxonomy and compatible reporting timelines for cybersecurity incidents. These requirements could draw on the serious incident reporting requirements in the EU’s AI Act or the NIS2 Directive’s requirement for a twenty-four-hour early warning and seventy-two-hour incident notification for covered entities. The United States’ Cyber Incident Reporting for Critical Infrastructure Act (CIRCIA) also provides similar reporting obligations for incidents within seventy-two hours. While these frameworks differ in triggers and entities covered, legislators should seek common definitions and compatible reporting procedures. Moreover, the reporting requirement should apply not just to the affected organization but also the model developer whose systems contributed to the incident.
- Establish designated points of contact: Both sides should establish by statute designated contacts within regulatory and law enforcement agencies for incident reporting and response if a security incident affects companies or organizations in the other’s jurisdiction. When a crisis strikes, a first step is often knowing whom to call.
- Work toward mutual recognition of government evaluator results: Congress and the European Parliament could direct the US Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) and the EU AI Office to coordinate their testing and evaluation benchmarks and work toward mutual recognition of their safety related benchmarks and testing. This would spare developers from having to undergo duplicate safety evaluations for release in major markets. Ideally, the United Kingdom’s AI Security Institute would be part of these efforts, as well.
- Trusted frontier model access program for defenders: When Hugging Face was attacked by a rogue swarm of OpenAI agents in July, it attempted to deploy a frontier model to defend itself, but that model’s guardrails prevented defenders from taking the necessary steps to stop the breach. The company ultimately turned to an open-weight Chinese model to stop the attack. To avoid inadvertently blocking defenders from responding to ongoing attacks, Congress and the European Parliament could build on existing trusted-access initiatives to give vetted cybersecurity defenders in the US and EU access to empowered versions of frontier models to better defend themselves from AI-enabled attacks.
- Baseline standards for third-party evaluators: After this summer’s spate of cybersecurity incidents, more US frontier labs have developed partnerships with independent evaluators as a voluntary safety measure. Legislators on both sides of the Atlantic could establish baseline accreditation criteria for third-party evaluators and standards for the minimum level of access labs must grant them. They should also establish guardrails to avoid conflicts of interest, including mandating financial independence and provisions periodically rotating evaluating teams.
Starting small
Even this phased agenda is ambitious and would certainly require more frequent engagement than one or two TLD sessions per year. I witnessed firsthand how disputes over conformity assessment for industrial goods dragged on during the Trade and Technology Council discussions under the Biden administration. That experience reinforces the case for starting with narrowly defined deliverables rather than attempting comprehensive alignment out of the gate.
An early deliverable for an initial AI-focused TLD session could be a joint set of recommendations on incident reporting and cross-border points of contact, followed by draft legislation in the House and a legislative initiative report in the European Parliament. Parliament could then request that the European Commission propose legislation that aligns with this shared policy framework.
For next steps, the TLD could further institutionalize AI-related exchanges between committees in both legislatures. The European Parliament’s TLD Steering Committee could assign a delegation from the AI Act Working Group between the IMCO and LIBE committees. From the US side, the delegation could draw from, if established, a renewed bipartisan group on AI policy, such as the proposed Select Committee on AI. Members of the Senate Commerce Committee could also be invited as observers or the US TLD delegation could be formalized and expanded to include Senate participants as Rep. Jim Costa (D-CA) proposed in 2022.
AI agent-driven cyberattacks pay no heed to jurisdictional boundaries. The developer may sit in California, the victim in Frankfurt, and the servers in Dublin. Researchers in France showed that recent attacks on internet-connected refrigerators across Europe could be replicated using Claude at a cost of about six dollars. Regulators on each side will learn of an attack on different timelines and under different definitions, or they may only read about it in the press after the fact. Next year offers a window to prevent that: a new Congress convenes, CISA’s long-delayed incident reporting rule is under White House review, and the AI Office is beginning to exercise its new powers for the first time.