The U.S.-Japan NEXT Alliance Initiative is a forum for bilateral dialogue, networking, and the development of joint recommendations involving a wide range of policy and technical specialists (in and out of government) to stimulate new alliance connections across foreign, security, and technology policy areas. Established by Sasakawa Peace Foundation USA with support from the Nippon Foundation, the goal is to help improve the alliance and how it serves shared interests, preparing it for emerging challenges within an increasingly complex and dynamic geostrategic environment. Launched in 2021, the Initiative includes two overlapping lines of effort: 1) Foreign & Security Policy, and 2) Technology & Innovation Connections.
Abstract
On June 10, 2026, Sasakawa Peace Foundation USA (through its U.S.-Japan NEXT Alliance Initiative) hosted a roundtable discussion on AI alignment challenges and national security, focusing on the distinct nature of these potential risks and the implications for U.S.-Japan alliance collaboration. As Caleb Withers and his co-authors argue in their recent working paper “Off Target: AI Alignment Challenges for National Security,” the binding constraint on AI potential for the future is most likely to be trust, rather than capabilities alone.[1] The U.S. and Japanese governments have important stakes in bolstering “alignment science” so that they are better able to assess misalignment risk and build trust for wider AI adoption. Withers presented their paper, which examines how AI misalignment creates unique national security risks that require specialized government attention, expertise, and evaluation infrastructure. Then, Dr. Amir Sadovnik from Oak Ridge National Laboratory outlined the lab’s work on AI security from both threat-to and threat-from AI perspectives, emphasizing the need for independent adversarial testing, domain-specific test beds, and understanding how fine-tuning affects model alignment. Follow-on group discussion among two dozen government and non-government experts from the U.S. and Japan explored international collaboration opportunities, with participants discussing the importance of government involvement, realistic test beds, data provenance challenges, and the need for harmonizing evaluation standards among allies. Key concerns raised included the speed of AI development outpacing policy formation, the role of government as both consumer and shaper of AI progress, and the balance between maintaining technical leadership while building trust with allies through information sharing.
Opening Remarks
Mr. Caleb Withers (Research Associate, Center for New American Security) opened the session by outlining their paper on AI alignment challenges (i.e., when an AI system pursues objectives other than those intended), arguing that frontier AI models like ChatGPT and Claude pose unique national security risks that require specific attention beyond general AI reliability or trustworthiness concerns. This is especially true, he said, given the speed at which AI is being developed and adopted, referencing the U.S. Department of Defense’s “AI Acceleration Strategy” that seeks to “integrate the bleeding edge of frontier AI capabilities across every mission area,” accepting that the risks of not moving fast enough outweigh the risks of imperfect alignment.[2] While perpetually perfect alignment in AI systems might not be achievable, the stakes are growing in this field as companies develop ever more capable products that are being deployed throughout the nation’s critical infrastructure, Anthropic’s restraint with its new cybersecurity agent Mythos (and subsequent U.S. export controls) notwithstanding.[3][4]
Although misalignment concerns with AI systems are not necessarily a new phenomenon, either caused by how models are trained or through possible data “poisoning” or other factors, newer frontier models have shown the capacity to recognize when they are being evaluated and have concealed aspects of their functions from developers to avoid poor marks. These challenges are not inherently catastrophic, Withers argued, if governments invest in the expertise and evaluation capacity to detect and manage them, but they do require dedicated research and assessment. Withers noted that two of the four pillars of the recently released U.S. National Security Memorandum 11 (on “Artificial Intelligence in the National Security Enterprise”) focus on “assurance” and “accountability.”[5]
In fact, beyond an emphasis on AI system reliability for government use, Memorandum 11 says that “the national security enterprise shall ensure, through contractual clauses or other means, that no commercial entity or adversary possesses the capability to prevent use of, disable or degrade, or materially modify without Federal Government knowledge and approval.” How to monitor this is left unspecified, and this is part of what Withers is calling attention to and encouraging tailored government investment. We know from experience that national efforts to bolster AI safety have benefited from international collaboration and information sharing, for example through collaboration among national AI safety institutes or consolidating incident reporting, and the same is likely to be true for emerging research in AI alignment science and performing evaluations of new AI systems under realistic conditions.
Withers recommended that governments invest in expertise, research capabilities, and evaluation infrastructure to assess AI alignment, noting that their impact could be enhanced through international coordination and developing technical methods to verify developers’ trust claims. He added that government procurement of AI systems should not necessarily penalize unorthodox “chain-of-thought” reasoning by AI models in government evaluations, lest it incentivize developers to hide such anomalies and then miss an opportunity to identify future misalignment risks. Overall, as Withers and his co-authors argue in their paper, “the government cannot afford to be a passive consumer of commercial progress. It must become a sophisticated customer and active shaper of the alignment landscape.”
Dr. Amir Sadovnik (Research Lead at the Center for AI Security Research (CAISER)) then introduced the work of CAISER at Oak Ridge National Laboratory, which he and Dr. Edmon Begoli founded just a few years ago as one of the first centers for AI security. Their work examines both threats to AI systems and possible threats from such systems, particularly from a government deployment and national security perspective. Focusing on threats-from AI, this can apply to potential adversaries using AI to undermine U.S. interests or to threats from the AI itself, and misalignment concerns are relevant to both cases. Sadovnik underscored the relative newness of this field of research, necessitated by the quick spread of AI technologies around the world, while related research funding streams and infrastructure struggle to keep pace.
Sadovnik also emphasized how complex the evaluation process can be, since AI models rarely operate in an isolated vacuum and are instead incorporated as part of a broader system including sensors, other software, and dynamic human interaction. Evaluating the system holistically is important, as is testing the system at scale and in a rapid manner to keep up with constant commercial innovation. Another aspect he highlighted was government use of small-form AI performing narrow tasks often at the edge, which is another area of research at their Center.
Their current research priorities driven by the national security enterprise include:
- Independent adversarial testing and evaluation at scale, with an emphasis on “independent” to provide objectivity apart from commercial-only analysis, by bombarding the model with different challenges in a safe environment under realistic conditions. Sadovnik introduced Oak Ridge’s Photon framework for red-teaming large AI models under conditions where the model is more likely to fail, in part by leveraging the Lab’s Frontier exascale supercomputer and a growing database of known attack methods.[6]
- Domain-specific test-beds and benchmarks that meet national security needs are another area of research, because the U.S. government has access to facilities and expertise that AI firms do not, such as with electricity grids, nuclear power plants, and other critical infrastructure environments.
- Understanding how fine tuning and the adoption of models (i.e., moving from a “base” model to a tailored model for a particular use) affects their alignment and existing guardrails. This is a relatively new area of research but growing in importance due to the possible evolution of risks related to misalignment or “jailbreak” potential.
- Developing methods to enhance AI explainability because it will become increasingly important to understand the national security implications of certain models as early as possible but to do so without stifling innovation. Sadovnik mentioned that cases have already been identified where apparent chain-of-thought output was not accurately reflecting what was going on within the model (for example, in cases of reinforcement learning toward a specific goal), so more research and improvement on this front will be key for addressing alignment challenges.
Sadovnik concluded with thoughts about international collaboration opportunities between the U.S. and Japan, calling attention to the Genesis Mission, which is a relatively new U.S. national initiative led by the Department of Energy to unite national supercomputers, massive datasets, and experimental facilities into a single AI-driven platform.[7] Japan became the first official international partner for the Genesis Mission in early June 2026, creating a bilateral initiative for research on AI science and other new technologies backed by a $500 million funding pledge from each side.[8] So, how to capitalize on this opportunity is clearly one important question for the alliance, as is the potential for sharing access to (and information from) domain-specific AI test beds in each country as a way to leverage these resources more economically for maximum benefit. Utilizing existing bilateral forums for knowledge sharing specific to AI misalignment research and related issues, among governments and with industry, is also important.
Discussion Summary
Group discussion featured questions for the opening speakers, as well as dialogue among the experts about how to invest in evaluation infrastructure, practical and useful ways to facilitate international collaboration, data quality and information sharing issues, and diplomatic and multilateral dimensions. Overall, the group seemed to advocate for a proactive but balanced view toward integrating misalignment into the underlying research and evaluation agenda, being forward looking, investing and accounting for some unique alignment aspects without losing sight of other important work on general safety, reliability, and cybersecurity.
While one private sector participant said frankly that misalignment rarely gets mentioned when his company thinks about utilizing AI products, which the company manages through a traditional cybersecurity lens of controlling “access” and “autonomy,” another researcher cautioned that alignment concerns will grow as agentic AI permeates a company’s operations. AI will become more empowered to made decisions about that control, and so alignment will at the very least become an increasingly important aspect of system reliability and security. The group also agreed that governments have an interest in being at the forefront of alignment science given the stakes involved for their own operations and for the national security enterprise and critical infrastructure.
A Japanese participant noted that his government recently approved funding to invest in new AI testbeds this year and is soliciting input from industry and the research community for how to spend these funds most effectively.[9] He asked how the U.S. is approaching these kinds of investments, and one U.S. researcher explained that testbeds are becoming increasingly domain specific, meaning that they are designed to evaluate, validate, and benchmark AI models for a specific industry or niche. This suggests that some coordination among governments investing in complementary testbeds might add value, if they can agree on what standards and benchmarks to use.
Another U.S. participant advocated for being ambitious about designing sufficient complexity and realism in the testing environment. Newer AI models can tell when they are being evaluated in a simplistic way, he argued, so real-world “sandboxing” is one possible approach, where relevant real-world data and components might be involved, albeit slightly isolated for secure testing to the point of failure. A third participant likened it to the early days of testing food-delivery robots in large cities when humans with clipboards would follow behind each robot throughout the exercise to monitor its behavior and reinforce safety.
A U.S. researcher emphasized that AI technology is developing much faster than any traditional policy making process can keep pace with, but to have at least some chance to provide meaningful oversight, governments need to invest in sufficient research and evaluation infrastructure to remain technically astute about the evolving technology. He added this point when the topic of Japan’s plan to double the size of its AI Safety Institute came up, and he encouraged Japan to empower the organization with talent, resources, and strong partnerships with Japanese laboratories and private firms. He explained that U.S. labs regularly engage in outreach and education for government agencies on the underlying AI components of software being used by government or by institutions they oversee that are related to national security or other “high-consequence” domains.
A Japanese research executive described his institute’s efforts to bolster its AI evaluation research capacity (AI for security and security for AI) with a view that improving the safety and reliability of AI systems is important for building trust and speeding up adoption. His institute is partnering with those from other countries (including the U.S.) on collaborative research projects and scientist exchanges, but the government is also encouraging organizations like his to strengthen sovereign AI capabilities. He asked how U.S. labs were balancing domestic priorities with international collaboration. A U.S. researcher explained that it can be difficult procedurally to work with foreign labs, and it is usually conducted most successfully when government agencies are involved and jointly commissioning the research in pursuit of shared national security goals. Despite the bureaucratic challenges, researchers from both countries emphasized the benefits of international partnering due in part to the cross-fertilization of ideas and different approaches to problem solving that takes place during such collaboration.
Another U.S. participant praised the work of the United Kingdom’s AI Security Institute as a good model to follow when investing in AI evaluation infrastructure, given its strong investment in testbeds and benchmarking and its frequent publishing of research results. She highlighted a useful recent UK AISI report on “loss of oversight” that benefited from evaluating chain-of-thought data. There will of course be times when testing and research results will need to remain private for national security reasons, but she advocated for as much public release as possible to share knowledge, help build trust, and make the evaluations themselves more interpretable. She suggested that some work be done to compare evaluations by different institutions with a goal of identifying best practices and enhancing their quality and usefulness. A different U.S. participant added that UK AISI was a good example of becoming excellent at something as a way to ensure a seat at the table when governments debate AI governance issues.
Fast-paced AI evolution is becoming a forcing function for nations to coordinate on AI governance issues including the topic of alignment, which used to be a largely theoretical concern but is becoming more concrete. Researchers are moving from abstract arguments on alignment to more specific applications with such things as drone behavior or cybersecurity agents. Although 2025 saw some international AI governance forums shift away from an emphasis on “safety” in favor of “impact,” the Claude Mythos release has refocused attention on AI safety. One U.S. participant suggested that we use this shift to introduce misalignment into these broader multilateral governance dialogues.
Participants from both countries said that there was an important role for the State Department/Foreign Ministry to play in terms of facilitating bilateral and international policy coordination on AI security issues. Allies need to develop and improve frameworks for sharing information that protect security as well as companies’ intellectual property. The Trump administration’s June 2 Executive Order on “Promoting Advanced AI Innovation and Security,” for example, directs the creation a government-led “cybersecurity clearing house” and “classified benchmarking process” to assess the cyber capabilities of “covered” models and help protect critical infrastructure by coordinating with “trusted partners.”[10] The Japanese government and certain Japanese firms will likely be among these trusted partners, as they were in the case of Claude Mythos, but a reliable process for this kind of coordination will need to be developed.[11]
An AI startup executive described a debate going on within the technical community about how soon the biggest AI safety, misalignment and related challenges will manifest, with firms like Anthropic being particularly vocal about near-term concerns, while others including Microsoft, Nvidia, Google and many in the applied AI startup community see no evidence that these problems are on the horizon. If governments overreact, he said, there was a danger of foregoing significant value creation and innovation. He suggested more of an “outcome-oriented” policy rather than preventative, and one way to address this could be to make the very largest closed-source models liable for mistakes or misalignment (given their higher stakes and available resources to research and mitigate) but have a lighter touch with others. Another participant agreed that incentivizing firms to invest adequately in safety is important, but this is still a shared challenge that is evolving quickly, so being empirical and forward-looking at the same time is important. Something that models can do expensively or with difficulty today will soon become commonplace, so there is a collective responsibility for safety and information sharing that should not be avoided.
A U.S. participant added the issue of data quality to the mix, noting that AI systems are increasingly generating data that is then consumed by other systems as part of their training, so this could be another aspect of collaborative research and information sharing among allies. A project like the Genesis Mission that uses primarily scientific data produced by government labs, agencies, or research universities should be easier to exert quality control, but there is little consistency in the commercial realm. Another participant pointed out that good data is not a guarantee for avoiding misalignment when we begin to let models learn and experiment on their own, so this is another area that we need to understand better.
Finally, on the military front, a U.S. participant worried that U.S. military adoption of AI could make it more difficult to engage with allies and partners on some of these AI implementation issues, depending on how the Pentagon resolves internal debates about balancing risks between AI safety and alignment versus the risks of slowing adoption. Interoperability with Japan’s Self-Defense Forces could become more complicated with non-disclosure requirements or impose new demands on Japan, and this will need to be coordinated with what the Japan side is comfortable with or prepared to handle.
A U.S. government participant emphasized the importance of cybersecurity in the military context, and this could be productive area of U.S.-Japan collaboration. AI and cybersecurity testing go hand-in-hand, and each of the U.S. military service branches maintains its own operational test agency that helps the Defense Department manage the volume and speed of testing demands with insights from warfighters. Relationships between U.S. and Japanese military testers of cybersecurity exist, but they can be strengthened and focused on linked systems. Cyber defense could become an important alliance mission as U.S. Forces Japan emerges as a new Joint Force Headquarters working closely with Japan’s Joint Operations Command. The U.S. participant recommended that the U.S. consider creating a Cyber Force as a new branch of military service that deals with the cyber aspect of military operations every day in a joint manner, and it could introduce a cyber operational test agency to focus on both offensive and defensive aspects.
NEXT Steps
The Genesis Mission presents an exciting new opportunity for U.S.-Japan science collaboration that can greatly benefit both countries and the world. However, it will be difficult to implement effectively due to such issues as information sharing protocols, data compatibility, communication challenges, budgeting procedures, aligning research priorities, and others. Examining this more closely for optimization will be one function of our quarterly roundtable going forward. Another area we want to focus on is fostering alliance dialogue on AI governance and security issues involving more private AI firms, so that we can help bridge gaps that might exist between the public and private sectors. It is critically important to enhance mutual understanding and a sense of common purpose among policy makers, research funding appropriators, scientists, and AI technical and business specialists in the private sector.
[1] Caleb Withers, Jay Kim, and Ethan Chiu, “Off Target: AI Alignment Challenges for National Security,” Center for a New American Security, March 24, 2026, https://www.cnas.org/publications/reports/off-target.
[2] “War Department Launches AI Acceleration Strategy to Secure American Military AI Dominance,” U.S. Department of Defense, Accessed June 2026, https://www.war.gov/News/Releases/Release/Article/4376420/war-department-launches-ai-acceleration-strategy-to-secure-american-military-ai/.
[3] “Claude Fable 5 and Claude Mythos 5,” Anthropic, Accessed June 2026, https://www.anthropic.com/news/claude-fable-5-mythos-5.
[4] Madhumita Murgia, George Hammond, Rafe Rosner-Uddin, and Joe Miller , “Anthropic scrambles after Trump administration freezes its top AI models,” Financial Times, June 14, 2026, https://www.ft.com/content/f6940d59-28f4-4ae4-a569-c6fc421e52b9?syn-25a6b1a6=1.
[5] “NATIONAL SECURITY PRESIDENTIAL MEMORANDUM/NSPM-11,” The White House, Accessed June 2026, https://www.whitehouse.gov/presidential-actions/2026/06/national-security-presidential-memorandum-nspm-11/.
[6] Audrey L Carcson, “Photon framework scales AI vulnerability discovery,” Oak Ridge National Laboratory, March 24, 2026, https://www.ornl.gov/news/photon-framework-scales-ai-vulnerability-discovery.
[7] Julia Medeiros Coad, “Genesis Mission’s AI-Driven Research To Strengthen US Grid, Supply Chains, and Technological Leadership,” National Laboratory of the Rockies, March 10, 2026, https://www.nlr.gov/news/detail/features/2026/genesis-mission’s-ai-driven-research-to-strengthen-us-grid-supply-chains-and-technological-leadership.
[8] “United States and Japan Announce Historic $1 Billion Partnership Under President Trump’s Genesis Mission,” U.S. Department of Energy, Accessed June 2026, https://www.energy.gov/articles/united-states-and-japan-announce-historic-1-billion-partnership-under-president-trumps
[9] “Japan targets $65bn in public-private physical AI investment by 2040,” Nikkei Asia, June 20, 2026, https://asia.nikkei.com/business/technology/artificial-intelligence/japan-targets-65bn-in-public-private-physical-ai-investment-by-2040.
[10] “PROMOTING ADVANCED ARTIFICIAL INTELLIGENCE INNOVATION AND SECURITY,” The White House, June 2, 2026, https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/.
[11] “Kazuaki Nagata, “Japanese government and banks granted access to Claude Mythos,” The Japan Times, June 3, 2026, https://www.japantimes.co.jp/business/2026/06/03/tech/japan-anthropic-mythos/.
