UN disarmament institute to launch first UN forum on military AI this week

What UNIDIR announced

The United Nations Institute for Disarmament Research, known as UNIDIR, will launch a forum on military uses of artificial intelligence at UN Headquarters in New York on Thursday, 22 October 2026. Its director, Robin Geiss, announced the plan at a press briefing on 20 October, saying the forum will be "the only multi-stakeholder body within the United Nations system that works on AI and international peace and security and in the military domain."

The forum belongs to UNIDIR's Centre of Excellence on AI, Peace and Security, which the institute set up in Geneva in June 2026. According to UNIDIR's event page, Thursday's side event to the General Assembly's First Committee, the committee that handles disarmament, is organised with Switzerland and Pakistan and opens expressions of interest for organisations that want to join. Speakers include Izumi Nakamitsu, the UN's High Representative for Disarmament Affairs, and the deputy UN ambassadors of Switzerland and Pakistan.

Geiss also announced three related projects to be rolled out over the same week: a framework for responsible behaviour by companies that build military AI, a shared glossary of AI security terms, and work on keeping AI capabilities out of the hands of non-state armed groups. UNIDIR's calendar lists separate events on the industry framework and the glossary on 23 October and on non-proliferation of AI on 21 October.

Why the UN institute says the timing matters

Geiss opened the briefing with a blunt question: "Are we going to let machines target and kill human beings?" According to the UN's own automatic transcript of the briefing, he said AI and autonomy are already "embedded widely across the military domain," from intelligence analysis and target identification to drones, cyber attacks and information operations, and that decision cycles are "compressing to machine speed," including in the nuclear domain.

His concern is less that a single weapon will go rogue than that human judgment is being squeezed out of the whole process. Asked how close the world is to full autonomy, he said it was hard to tell because AI is "creeping in" to early warning, reconnaissance and information analysis, but that the process "is clearly on track to crowding out human decision making." Militaries, he added, have no wish to lose control; the risk comes from an arms race for speed, with shorter times between finding and striking a target and a growing number of targets engaged each day.

Geiss argued that concern is now crossing geopolitical lines. UN News notes that in September the United States proposed an AI notification mechanism with China so the two countries could warn each other about AI incidents threatening national security. He also pointed to the Secretary-General's calls for a legally binding treaty on lethal autonomous weapon systems. National militaries are moving the other way at the same time: the US Army has picked a vendor to build ECHOTRIBBLE, an autonomous hacking agent meant to work with minimal human oversight.

A rulebook for defence technology companies

The first new project targets the companies that design and build military AI. Geiss said choices about system design, training data, testing and support after deployment decide whether AI-enabled military capabilities are reliable, predictable and lawful, and that this stage is largely in industry's hands.

UNIDIR is therefore drafting a framework of responsible industry behaviour: practical, voluntary guidelines grounded in international law and the UN Guiding Principles on Business and Human Rights. He said it should set a common baseline across the technology life cycle and go on to inform government procurement and public-private partnerships. Consultations launching this week are meant to include large and emerging defence technology firms, software developers, dual-use service providers, subcontractors and investors, and UNIDIR is working on it with the Office of the High Commissioner for Human Rights.

The guidelines would be voluntary, but they arrive as the line between commercial AI labs and defence suppliers becomes contested. In the United States, Anthropic is still fighting a Pentagon decision to label it a supply-chain risk after it refused to drop limits on weapons use, which shows how much a company's own usage rules can matter to how its models are used in war.

A shared vocabulary for autonomy and safety

The second project is about language. Geiss said states often use words such as autonomy, safety and security to mean very different things, and that the ambiguity seeps into national positions and makes agreement and trust harder to build.

UNIDIR plans to launch work on an "AI security lexicon" on Friday. It is meant both to clarify key concepts and to record where understandings still diverge, so that states can at least see where they disagree. The goal, Geiss said, is to help countries "find common ground and to engage on an equal footing" on AI in the military domain. A glossary sounds modest, but long-running talks on lethal autonomous weapons have repeatedly stalled over definitions, such as what counts as meaningful human control.

Keeping AI capabilities away from armed groups

The third project deals with misuse of AI by non-state armed groups, terrorist organisations and criminal networks. UNIDIR is convening a discussion on non-proliferation of AI with UNITAR, the UN's training and research institute, in Austria this week. Geiss said emerging research shows such groups already using AI for operational planning, execution and logistics, a risk he said has received little attention.

He argued that existing non-proliferation regimes and export controls provide a starting point but were not built for AI, and that the aim is to find gaps and propose measures at national and international level without blocking peaceful uses. Cases such as the ARTEX agent that GMO's security unit found attacking Japanese companies show how quickly capable tools spread beyond states. Asked about biological risks, Geiss called the chance that powerful language models could reveal biosecurity knowledge that was previously out of reach "a very serious concern" and a core focus of UNIDIR's research.

Who pays and who takes part

UNIDIR is funded by voluntary contributions. Geiss said the centre has start-up funding to carry it through its first five years, and that all five permanent members of the Security Council, the United States, Russia, China, the United Kingdom and France, are donors to the institute for different parts of its work.

On industry, he said some companies have already joined the process and others prefer not to be named. Asked for examples, he named Microsoft, which he said is a UNIDIR donor and will take part, and Hitachi America, and said a list of companies that agreed to be listed is public. These remarks come from the UN's automatic transcript, which the UN says is not an official record, so names and wording may contain transcription errors.

Over the coming months the centre plans to help countries draft national approaches to AI in security and defence and to support Switzerland ahead of the Geneva AI Summit in June 2027, where Geiss said military AI "will certainly no longer be allowed to linger in the shadows."

What the forum cannot do

None of this is binding. The forum is a space for exchange and consultation, the industry framework will be voluntary, and the lexicon is a reference document. UNIDIR is a research institute, not a negotiating body, so any treaty on autonomous weapons would still have to come from states, through the First Committee or the Convention on Certain Conventional Weapons talks in Geneva.

Geiss himself acknowledged that geopolitical competition is a "serious obstacle" and that expectations must be managed. UNIDIR has not yet published the draft framework, the membership of the forum or a timeline for either. Industry participation is so far partial and partly anonymous. Domestic debates over AI safety rules, like the UK parliamentary hearing where OpenAI and Anthropic backed binding rules, have so far focused on civilian models rather than military systems.

Why it matters

Until now the UN has had no standing body where governments, companies, researchers and civil society discuss military AI together; talks have been split between disarmament forums focused narrowly on autonomous weapons and broad AI governance processes that avoid defence. A permanent forum with guaranteed funding for five years gives the topic a fixed home.

For AI companies, the industry framework is the part to watch. If it becomes a reference for government procurement, as Geiss hopes, design and testing choices made in labs could become criteria for selling to militaries, even without a treaty.