Hello, this is Hamamoto from TIMEWELL. When people hear the word AI, many of them picture peaceful uses such as writing text or generating images. Yet the same technology is quietly moving to the front line of the battlefield and national security. AI that selects targets, weapons that may decide on their own whether to attack, and a move by governments to draw a line that says "this AI model must not be handed to other countries." All of these look like a distant world, but they are in fact connected to the foundations of the AI services we use every day.
In June 2026, U.S. Eastern Time, the U.S. government was reported to have issued a directive to major AI companies to suspend foreign-national users' access to their top-tier AI models. What was stopped was not a "weapon" but "AI itself."
In this article, I will trace how AI is used in the military, the international rulemaking around lethal autonomous weapons, and the trend of AI models becoming subject to state control, breaking down the jargon as much as I can along the way.
What it means for AI to become a weapon
First, I want to clear up a misunderstanding. What is actually advancing right now is not "robot soldiers like the Terminator." It is a far more mundane and far more widely used form of AI. It is AI as software that reads through the enormous volumes of data gathered from the battlefield and sorts out where things are and who might become a target.
A symbolic example is Project Maven, which the U.S. Department of Defense launched in 2017[^1]. This was an effort to use machine learning to automatically analyze ISR footage, that is, intelligence, surveillance, and reconnaissance imagery, collected by drones and sensors. Picture AI taking over the grunt work that a human would otherwise do by staring at a monitor for hours looking for vehicles and people. According to reports, the "Maven Smart System" in this lineage has been used in Ukraine since 2022 to pinpoint the location of Russian equipment. The director of the U.S. National Geospatial-Intelligence Agency (NGA) reportedly said in September 2025 that Maven would begin delivering "machine-generated" intelligence at scale to combatant commanders by mid-2026, but this is a citation by way of a secondary source, and the original statement remains unconfirmed at this time.
Another case that drew global attention was target selection in Gaza. Multiple media outlets reported that the Israeli military generated human target candidates with a system called "Lavender" and selected physical targets such as buildings with "The Gospel (Habsora)"[^2]. Vivid figures have been reported as well, such as target candidates reaching tens of thousands and human verification taking only a very short time, but these are descriptions based on reporting, and the Israeli military has denied their content. They cannot be asserted as fact.
To state my own stance, what truly frightens me here is not so much that "AI fires on its own," but that "humans accept what AI has selected without scrutinizing it enough." The better AI becomes as a sorting clerk, the more humans want to delegate judgment to it. That very slackening connects directly to the debate over lethal autonomous weapons that I turn to next.
Replace siloed classification work with AI.
METI's FY2024 data shows 52% of foreign exchange law violations stem from classification errors. TRAFEED cuts determination time by ~70% and stores structured rationale for every decision.
Building international rules around lethal autonomous weapons (LAWS)
Lethal autonomous weapon systems are called LAWS for short. The rough image is of weapons that find and attack targets on their own, without a human being meaningfully involved in the decision to pull the trigger. That said, there is still no single internationally agreed definition of "where autonomy begins" or "how much human involvement makes it permissible." Hammering that out among states is exactly where we are now.
The main stage for the debate is the Group of Governmental Experts (GGE) under the Convention on Certain Conventional Weapons (CCW), which meets in Geneva. According to reporting and materials published by international bodies, since 2023 this group has been tasked with building "elements of an instrument" by consensus, in a way that does not prejudge the premise of regulation, and its final report is expected to be submitted to the CCW Review Conference convening in November 2026[^3]. The 2026 sessions are reported to be scheduled in Geneva for early March and from late August into early September, for a total of about ten days. What I want to flag is that this is not the stage where "a treaty has been concluded"; it is still the stage of continuing to lay the groundwork for negotiations. The UN Secretary-General and the President of the International Committee of the Red Cross (ICRC) are said to be calling on states to conclude a legally binding instrument by the end of 2026, but a call and an agreement are two different things.
There is movement on the UN General Assembly side as well. According to UN meeting records, the First Committee of the General Assembly adopted a resolution on autonomous weapons in November 2025 for the third year in a row[^4]. This was adopted in the plenary as Resolution 80/57, and the recorded vote in the UN's official records is published as 164 in favor and 6 against. To add one caveat here, the number in favor shifts depending on which stage and which source you look at: the committee-stage vote and tallies by civil society groups have been reported at around 156[^5]. When consulting primary sources, it is safest to use the recorded plenary vote, which is the confirmed figure, as your baseline. Those reported to have voted against were Belarus, Burundi, North Korea, Israel, Russia, and the United States.
The structure is one in which a large number of states favoring a ban face off against a few heavily armed states that are cautious about regulation. According to civil society groups and reporting, at the September 2025 session Brazil took the lead and 39 states parties plus several observer states issued a joint statement that they were "ready to begin negotiations," while the actual start of negotiations has been blocked by countries such as Russia, Israel, India, Australia, South Korea, and the United States[^6][^7]. Rulemaking appears to be progressing, but honestly, where we stand now is stalled just short of the crucial matter of "binding force."
How Japan approaches LAWS
So what is Japan's position? According to the Ministry of Foreign Affairs, Japan states that it "will not develop fully autonomous lethal weapons that operate beyond the reach of human involvement," and that it will not conduct research and development of equipment whose use is not permitted under international or domestic law[^8]. The key phrase is "significant human involvement." Japan is said to take the position that such significant human involvement is indispensable to the use of weapons, and that regulation should be narrowed to fully autonomous systems that lack that involvement.
Put the other way around, the idea is that autonomous systems in which humans are properly involved should not all be uniformly restricted. According to the Ministry of Foreign Affairs, autonomous technology with assured human judgment has security significance in reducing human error and in carrying out defense with fewer personnel. For Japan, where the falling birth rate makes it hard to secure Self-Defense Force members, easing manpower burdens is a pressing theme as well. Japan has submitted working papers laying out its thinking to the UN and the CCW, and the most recent submission is reported to be in May 2024.
This is a point where I want to choose my words carefully. I was not able to directly verify the relevant Ministry of Foreign Affairs page this time, and the above reflects search-result snippets cross-checked against multiple secondary sources. I would ask you to read the finer points of the wording on the premise that they should be confirmed against the primary source.
For my part, I feel that this line drawn around "significant human involvement" is realistic. The idea of banning only fully autonomous weapons while permitting the scope for which humans bear responsibility can serve as a landing point that puts a brake on the technology without rejecting it outright. The hard part, though, is the substance of what "significant" actually means. As became an issue in the Gaza reporting, even if a human is nominally verifying, if that verification is no more than a few seconds of rubber-stamping, the involvement is in name only. To make the principle Japan upholds truly meaningful, I believe we need a deeper layer of discussion about how to guarantee the "quality" of involvement.
When the AI model itself became subject to state control
From here is the point that is the newest of the past year or two and yet tends to be overlooked. Until now, security interest has been directed at "things" such as missiles, fighter jets, semiconductors, and manufacturing equipment, and at the "technology" of their design information. But over 2025 and 2026, the United States began directly managing AI models themselves as security assets.
The first foreshadowing came in export control. According to commentary from law firms and others, in January 2025 the U.S. Bureau of Industry and Security (BIS) published an interim final rule that, for the first time in history, added AI model weights (the massive cluster of parameters obtained through training, which amounts to the very contents of the model) to the scope of export control[^9]. A new classification number, ECCN 4E091, was established, and it came to be known as the "AI Diffusion Rule." Yet the policy reversed before the rule took effect; according to reporting, the Commerce Department under the Trump administration formally rescinded this rule in May 2025[^10]. Whether or not to make AI models subject to control: the U.S. government itself wavered on that judgment.
What did not end with the rescission is 2026. According to reporting, in July 2025 the U.S. Department of Defense awarded contracts of up to USD 200 million each to Anthropic, Google, OpenAI, and xAI, accelerating the military use of frontier AI. Meanwhile, in February 2026, in a statement issued in the name of CEO Dario Amodei, Anthropic made clear that it would not allow its technology to be used for two purposes: fully autonomous weapons and large-scale domestic surveillance[^11]. The company also stated that "today's frontier AI is not reliable enough to be entrusted with fully autonomous weapons." In response, at the end of February 2026 the U.S. administration was reported to have directed federal agencies to stop using Anthropic's technology, and almost simultaneously OpenAI announced a deployment contract with the Department of Defense. OpenAI is said to have reached an accommodation under a framework of "lawful use," while prohibiting two things: large-scale domestic surveillance and lethal autonomy without human approval[^12]. The difference between Anthropic, which made its red lines contractually inviolable, and OpenAI, which agreed via a commitment to legal compliance, stood out clearly here.
And the decisive blow came in June 2026. According to reporting, the U.S. government ordered Anthropic to suspend its top-tier models Fable 5 and Mythos 5 for all foreign nationals[^13]. Anthropic, saying it had no means of screening foreign nationals in real time, is reported to have immediately disabled both models for all customers. The reason given is that a technique had come to light for working around Fable 5's safeguards to access the cyber capabilities of Mythos, which is adept at discovering serious software vulnerabilities. Furthermore, at the end of the same month, OpenAI announced its new model GPT-5.6 but was reported to have narrowed its provision to around 20 "trusted partners" at the U.S. government's request[^14]. Reporting frames this as the first instance of the U.S. government preemptively asking an AI company to restrict provision before release. As an article that digs into this whole sequence more from the angle of export control, I think reading The era when AI models become subject to export control alongside this one will help connect the dots into a line.
What is being asked of Japanese companies as users and builders
Having read this far, you might feel that "this is a story about American military and politics, and has nothing to do with Japanese business." But I believe Japanese companies in particular should take this on as their own concern. There are two reasons.
One is the risk on the user side. Many Japanese companies build U.S.-origin AI and cloud services into their operations. If that model is suddenly cut off one day for security reasons, the operations running on top of it stop along with it. The suspension of foreign-national access in June 2026 spread to all customers because Anthropic could not sort out foreign nationals. In other words, "we are a Japanese company, so we are out of scope" is not necessarily true. Which AI, and which country's technology, do you depend on? Is there an alternative when it stops? This inventory is becoming part of business continuity planning going forward. The point that access itself becomes subject to control is laid out from a practical perspective in How to prepare for the era when the AI you use suddenly stops one day.
The other is the risk on the builder side, the provider side. Once AI models and their related technology come to be classified as subject to export control, you have to judge, case by case, who you may provide what to. In the world of export control, the work of confirming whether a given technology or product falls on a regulatory list is called classification screening. Furthermore, there is a concept called "deemed export," under which letting foreign-national employees or overseas offices use technology is regarded as an export even without crossing a border. Once access to AI models or APIs comes within this scope, managing who inside the company may touch which AI, and how far you let business partners use it, becomes far more complex all at once.
This is where the export control AI agent we at TIMEWELL developed, TRAFEED (formerly ZEROCK ExCHECK), can prove useful. TRAFEED is a service we built to support classification screening of whether a product or technology falls under regulation, screening of counterparties, and monitoring of regulatory updates. The more military use and state control of AI advance, the more we need a mechanism to continually confirm, for AI models and related technology as well, "is this OK to hand over?" and "is the counterparty problem-free?" It replaces judgments that rely on a staff member's memory or intuition with operations aligned to the latest regulations. It is unglamorous, but this is a realistic step toward leveling the risk on both the user side and the builder side.
The lines that will be drawn next, and how companies can prepare
Finally, let me organize the points to come. LAWS regulation (the CCW and UN debates) and state control over corporate AI models are, as systems, separate tracks. The former is the disarmament context of constraining how weapons are used; the latter is the management context of who AI is handed to as a security asset. Yet the two are clearly connected by the questions of "how much significant human involvement to require," "what lawful use means," and "where to draw the line that must not be crossed, such as fully autonomous weapons and large-scale surveillance." The exchange between Anthropic and the U.S. government was an event in which that connection point erupted in the form of a corporate red line.
Seen from the technology side, this is also a repeat of a pattern we have seen before. One tech-focused outlet characterized the suspension of Fable 5 and Mythos as the first real-world test of whether frontier AI can truly be contained through export control, overlaying it with the history of how attempts to suppress encryption technology and spyware through export control once failed to function adequately[^15]. Digital technology can be copied and crosses borders easily. AI model weights are no exception. Within voluntary frameworks like the Wassenaar Arrangement, which brings transparency among states to exports of conventional arms and dual-use technologies, whether AI weights and training data can be treated as "technology" or "software" is something that will be tested in earnest from here on[^16].
What can companies do? There are no flashy answers. First, grasp the AI you depend on and its sources of supply, and estimate the impact and the alternatives if it stops. Second, inspect whether anything among the technology or AI access you provide could fall within the scope of regulation. Third, accept that regulations and entity lists are moving targets, and review your judgments continuously rather than treating them as a one-time exercise. These three are not the conclusion to the grand theme of military AI, but rather practical preparations you can begin tackling tomorrow.
If you want to work out what becomes a risk in your own case and where to start, please feel free to reach out via an individual consultation. At the intersection of AI and security, we can start by taking inventory of your export control and economic security responses together. Precisely because we live in an era when AI can become both a weapon and an object of control, this is a time to get ready, early, to put into words "who are we handing what to."
References
[^1]: Project Maven — Wikipedia — updated periodically — https://en.wikipedia.org/wiki/Project_Maven [^2]: AI-assisted targeting in the Gaza Strip — Wikipedia — updated periodically — https://en.wikipedia.org/wiki/AI-assisted_targeting_in_the_Gaza_Strip [^3]: Lethal Autonomous Weapon Systems — UNODA (UN Office for Disarmament Affairs) — updated periodically — https://disarmament.unoda.org/en/our-work/emerging-challenges/lethal-autonomous-weapon-systems [^4]: GA Adopts More Than 60 Resolutions, First Committee — UN Meetings Coverage — 2025-11 — https://press.un.org/en/2025/ga12736.doc.htm [^5]: 156 States Support UNGA Resolution — Stop Killer Robots — 2025 — https://www.stopkillerrobots.org/news/156-states-support-unga-resolution/ [^6]: September 2025 GGE Joint statement — Stop Killer Robots — 2025-09 — https://www.stopkillerrobots.org/news/september-2025-gge-joint-statement/ [^7]: Geopolitics and the Regulation of Autonomous Weapons Systems — Arms Control Association — 2025-01 — https://www.armscontrol.org/act/2025-01/features/geopolitics-and-regulation-autonomous-weapons-systems [^8]: Lethal Autonomous Weapon Systems (LAWS) — Ministry of Foreign Affairs of Japan — last updated unconfirmed — https://www.mofa.go.jp/mofaj/dns/ca/page24_001191.html [^9]: New U.S. Export Controls on Advanced Computing Items and AI Model Weights — Sidley Austin — 2025-01 — https://www.sidley.com/en/insights/newsupdates/2025/01/new-us-export-controls-on-advanced-computing-items-and-artificial-intelligence-model-weights [^10]: The US AI Diffusion Rule (rescission background) — United States Studies Centre — 2025 — https://www.ussc.edu.au/the-us-ai-diffusion-rule [^11]: Statement from Dario Amodei on our discussions with the Department of War — Anthropic — 2026-02-26 — https://www.anthropic.com/news/statement-department-of-war [^12]: OpenAI announces Pentagon deal after Trump bans Anthropic — NPR — 2026-02-27 — https://www.npr.org/2026/02/27/nx-s1-5729118/trump-anthropic-pentagon-openai-ai-weapons-ban [^13]: Anthropic Says US Orders Halt to Foreign Access for Fable 5, Mythos 5 — Bloomberg — 2026-06-13 — https://www.bloomberg.com/news/articles/2026-06-13/anthropic-says-us-limits-foreign-access-to-fable-5-mythos-5 [^14]: OpenAI limits GPT-5.6 rollout after government request — TechCrunch — 2026-06-26 — https://techcrunch.com/2026/06/26/openai-limits-gpt-5-6-rollout-after-government-request-says-restrictions-shouldnt-be-the-norm/ [^15]: From PGP to Mythos, history shows why cyber export control doesn't work — TechCrunch — 2026-06-19 — https://techcrunch.com/2026/06/19/encryption-spyware-and-now-mythos-history-shows-why-cyber-export-control-doesnt-work/ [^16]: Understanding U.S. Allies' Current Legal Authority to Implement AI and Semiconductor Export Controls — CSIS — 2025 — https://www.csis.org/analysis/understanding-us-allies-current-legal-authority-implement-ai-and-semiconductor-export