NATO Is Burning 300 Billion AI Tokens a Month: Inside OpenAI’s Daybreak Cyber Defence Deal

NATO has quietly become one of the biggest users of artificial intelligence on the planet — at least by one measure. OpenAI said on November 17, 2026, that the alliance is using up to 300 billion input and output tokens a month through its cybersecurity models, as NATO’s core institutions put frontier AI to work hunting software flaws before attackers can exploit them, as reported by The Hindu BusinessLine.

The arrangement runs through the NATO Communications and Information Agency (NCIA), the body that delivers technology and cybersecurity services to the alliance’s core institutions. It is the clearest sign yet that AI has moved from the briefing room to the front line of national cyber defence — and it comes with real results, real questions, and a race against attackers who are using the same technology.

AI magnifying glass scanning lines of computer code and highlighting security vulnerabilities

What the NATO–OpenAI Deal Actually Covers

OpenAI is giving NATO’s core institutions access to advanced cybersecurity models for authorised defensive work only. That work covers four jobs: reviewing software, investigating suspicious activity, checking whether suspected weaknesses are real, and developing and testing fixes.

Access is not automatic. OpenAI reviews each request, and approved users may only work on systems they own, operate, or are explicitly authorised to test. The models are for defenders, not for offensive operations — a distinction OpenAI has been careful to draw as its government work expands.

The Defense News confirmed the partnership runs through the NCIA under the Daybreak programme.

300 Billion Tokens: What the Number Really Means

Three hundred billion tokens a month is an enormous amount of AI processing — roughly the text of millions of books, read and written every month. But the figure needs careful reading: it measures how much work the models are doing, not how many attacks they stopped or flaws they found. OpenAI has not disclosed the financial terms of the arrangement, named the specific models NATO uses, or said when the cooperation began.

Even so, the scale tells its own story. NATO is not running a pilot. This is production-scale AI defence, and the alliance is studying the usage data closely. NCIA General Manager Dr Dylan Browne said the experience would help NATO understand token consumption, usage patterns, and the economics of AI models, supporting future cost optimisation and procurement decisions — in other words, this deal will shape how the alliance buys AI security tools for years.

Daybreak: OpenAI’s Programme for Trusted Defenders

The cooperation runs through Daybreak, OpenAI’s programme for accredited institutions and the US government, which launched in May 2026. Daybreak channels advanced models — combined with the agent features of OpenAI’s Codex coding tool — to authorised cyber defenders doing approved security work.

AI coding agents are moving fast, and Daybreak pairs those agent capabilities with OpenAI’s models.

“AI-powered cyber defence models represent a generational leap in cyber security. They accelerate vulnerability discovery, detect emerging threats at machine speed and enable defenders to stay ahead of adversaries,” said Dr Dylan Browne. Sasha Baker, OpenAI’s Head of National Security Policy, put it more bluntly: “Through Daybreak, we’re putting more capable tools in [NATO defenders’] hands to help them find and fix vulnerabilities before attackers can exploit them.”

Superpower Daily notes Daybreak launched in May 2026 for accredited institutions and the US government.

The European Results So Far: ENISA, CERT-EU and CERT Polska

Beyond NATO itself, OpenAI has extended the same access to the European Union Agency for Cybersecurity (ENISA) and national cyber agencies in France, Germany, and Poland — and the results are already public. ENISA used the models to identify four vulnerabilities in software used across EU institutions, with fixes published for all four.

A joint report by ENISA and CERT-EU, covering 27 software projects, found that 95 per cent of 88 potential issues assessed by software maintainers across the models tested were confirmed as genuine. Poland’s CERT Polska used the models to help identify six vulnerabilities in MikroTik RouterOS; the findings were tested by researchers, disclosed to the vendor, and fixes were released. These are concrete, verified wins — not benchmarks on a slide.

Europe is also building sovereign, open AI models in Europe to reduce dependence on US providers.

Glowing digital shield of network connections protecting a map of Europe

Tested in Battle: Crossed Swords 2026

The partnership was also put through its paces in the 2026 Crossed Swords exercise, organised by the NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE) in Tallinn, Estonia, from November 2 to 6. Participants built, deployed, and evaluated AI agents for cyber operations, with a strong emphasis on maintaining human oversight.

It was the first time Daybreak was used in the exercise’s technical environment — a rehearsal for a future in which AI agents do much of the routine work of cyber defence while humans keep control of the critical decisions. OpenAI executives discussed the announcement at NATO Edge in Izmir, Turkey, underlining how central the alliance has become to the company’s government strategy.

The Other Side of the Coin: AI Is Arming Attackers Too

There is an uncomfortable symmetry to this story. The same class of models that NATO is using to find flaws is being used by attackers to find them first. Google’s threat intelligence team warned this week that AI agents are already turning patches into exploits, with fully autonomous attacks moving from theory to reality.

That is why the race matters. Defenders have to find and fix a vulnerability before an attacker finds and exploits it — and AI is compressing that window from weeks to hours. Microsoft’s new cyber-specific AI model, unveiled at Ignite this week alongside GPT-6.2 Astra, is aimed at exactly the same problem from the vendor side.

What This Means for Cyber Defence — and for the Rest of Us

NATO’s 300-billion-token habit is a milestone, not a finish line. It proves that frontier AI models can do useful, verifiable defensive work at scale — finding real flaws, in real software, that real vendors then fixed. It also shows how fast the market is moving: governments are now among the biggest customers for AI, and their buying decisions will shape which security tools get built next.

For everyone else, the lesson is simpler. The software that runs hospitals, power grids, banks, and government services is now being defended — and attacked — by AI. The side that finds the flaw first wins, and for the first time, the defenders have tools as fast as the attackers. Expect more governments to follow NATO’s lead, and expect the token counts to keep climbing.

AI in cyber defence has crossed from promise to practice. NATO’s deal with OpenAI is the biggest proof point yet — 300 billion tokens a month, real vulnerabilities found and fixed, and a procurement signal that will echo across the industry. The question now is not whether AI belongs in cyber defence, but how quickly every other defender can catch up.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top