OpenAI said on Tuesday, Nov. 17, that it works with the NATO Communications and Information Agency, the alliance's technology and cyber agency, to give NATO's core institutions access to its advanced cybersecurity AI. The aim, OpenAI says, is to help protect the alliance's networks. The access comes through Daybreak, OpenAI's program that offers AI to what it calls trusted defenders. OpenAI says the program is for approved security work: reviewing software, looking into suspicious activity, confirming vulnerabilities and building and testing fixes. A vulnerability is a weakness in software or a system that an attacker could use.
OpenAI says NATO has used up to 300 billion tokens a month in this work. Tokens are the small chunks of text, often parts of words, that AI models read and write. The count covers both what NATO sends the AI and what the AI writes back. For comparison, the EU review described below used about 611 million tokens in total. Dylan Browne, who runs the NATO agency, called AI cyber defense models "a generational leap in cyber security." He said that as NATO uses more of them, it will learn more about how much it uses and what the models cost to run, which "will help us optimise costs, evaluate competing solutions and make informed procurement decisions."
OpenAI also supported Crossed Swords 2026, a training exercise that NATO's Cooperative Cyber Defence Centre of Excellence ran in Tallinn, Estonia, from Nov. 2 to 6. The centre says the exercise brought together more than 250 people from over 40 countries, NATO members and others. The exercise trains military cyber teams to attack as well as defend, timed to support military operations. OpenAI was one of more than 20 outside partners. A new part of the exercise had participants build and test AI agents for cyber operations and then work alongside those agents, in training that stressed keeping people in charge. An AI agent is AI that can carry out tasks on its own, such as using apps and websites, instead of only answering questions. It was the first time the exercise used OpenAI's Daybreak. It also used AI that ran on the exercise's own computers.
OpenAI says it has also given the EU's cybersecurity agency, ENISA, and national cyber agencies in France, Germany and Poland access to its cyber models. A report ENISA published on Nov. 12 with CERT-EU, the EU institutions' own cybersecurity service, gives an outside look at how well tools like these work. In August, ENISA used several OpenAI models and Anthropic's rival model Mythos 5 to read the code of 27 software projects that EU bodies maintain. The AI read only the code and never tested live systems. The models flagged 220 possible problems, and the report gives results for all the models together, not for each company.
By the time the report came out, the projects' owners had checked 88 of the 220, about two in five, and confirmed 95% of those as real problems. The rest were still being checked. ENISA found the models tended to rate problems as slightly more serious than the owners did. ENISA warns that the number of findings can put a heavy load on the people who must check them. It says AI can speed up such reviews but "cannot replace human judgement." ENISA says it used the models free of charge.
OpenAI calls this NATO's early use of its tools, but it has not said when that use began or which models NATO uses, and neither side has said whether NATO pays. Nor has NATO said what the tools have found in its own networks.