NATO Institutions Now Burn 300 Billion OpenAI Tokens a Month
OpenAI's Daybreak program supplies NATO and European cyber agencies with frontier models. ENISA found four EU-wide vulnerabilities; CERT Polska flagged six in MikroTik RouterOS.
By Daniel Okafor
5 min read

What's News
- NATO institutions consume up to 300 billion raw input and output tokens per month under the Daybreak program.
- ENISA used OpenAI models to identify four vulnerabilities in software used across EU institutions; all four were fixed.
- A joint ENISA and CERT-EU November report found 95% of 88 assessed potential issues were confirmed as genuine.
- CERT Polska used OpenAI models to help identify six vulnerabilities in MikroTik RouterOS, with vendor fixes verified.
- OpenAI has not disclosed financial terms, specific models, or liability provisions for the NATO arrangement.
NATO institutions now consume up to 300 billion raw input and output tokens per month through OpenAI models. That figure, published in OpenAI's own announcement of its "Daybreak" program, is the clearest measure yet of how deeply frontier AI has moved into allied cyber defense work.
OpenAI has formally detailed its expanding work with NATO and European national cyber agencies. The company positions Daybreak as a vehicle to put frontier AI models into the hands of authorized defenders for three tasks: vulnerability discovery, fix validation, and operational cyber defense. The announcement names NATO's Communications and Information Agency (NCIA) as the institutional anchor of the arrangement on the Alliance side.
The numbers involved are large by any standard of government AI adoption. Up to 300 billion tokens per month of raw input and output represents sustained, heavy usage across Alliance institutions, not a pilot. OpenAI also supported the 2026 Crossed Swords exercise, organized by the NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE). The exercise added a new training stream: participants built, deployed, and evaluated AI agents for cyber operations using Daybreak inside the technical environment.
What does NATO get out of the partnership?
NATO called the capability a "generational leap in cyber security" — one that accelerates vulnerability discovery and lets defenders stay ahead of adversaries at machine speed. That phrasing, from the Alliance itself, signals how NATO frames the stakes: speed as the decisive variable in modern cyber defense.
The Alliance also has a hard-nosed procurement motive. NATO said it will use the partnership to build a deeper understanding of token consumption, usage patterns, and the underlying economics of AI models to inform future procurement decisions. In plain terms, the Alliance is learning what frontier AI costs at institutional scale before it commits to larger buys.
Sasha Baker, OpenAI's Head of National Security Policy, framed the effort in political terms. The goal, she said, is ensuring "the people defending our democracies should have the most advanced technology on their side." OpenAI said it remains fully committed to NATO's principles for responsible AI adoption and to delivering solutions that are safe and secure. The company emphasized maintaining human oversight during the Crossed Swords training.
Which European agencies already have access?
Access extends well beyond NATO headquarters. OpenAI said it has already provided its cyber models to defenders across Europe, including:
- ENISA, the EU's cyber agency
- ANSSI, France's national cyber agency
- Germany's national cyber agency
- Poland's national cyber agency
- Ukraine's Ministry of Digital Transformation
The results are concrete. ENISA reported using the models to identify four vulnerabilities in software used across EU institutions, with fixes published for all four.
A joint November report from ENISA and CERT-EU described a review of 27 software projects using frontier models — including OpenAI models accessed through Daybreak — plus tools such as the Codex security plug-in. The verification rate matters most: across the models tested, 95% of the 88 potential issues assessed by software maintainers were confirmed as genuine. That ratio addresses the core criticism of AI-assisted vulnerability research, that models produce noise. Here, almost every flagged issue held up under human review.
Poland's results follow the same pattern. CERT Polska used OpenAI models to help identify six vulnerabilities in MikroTik RouterOS, coordinated disclosure with the vendor, and confirmed that MikroTik's released fixes prevented the attacks researchers had observed. The lifecycle ran end to end: discovery, vendor coordination, patch, and validation.
Ukraine's Ministry of Digital Transformation is using Daybreak to strengthen the cyber defense of civilian infrastructure. Its stated goals are identifying software vulnerabilities and developing and testing fixes more quickly amid persistent cyber attacks.
Why is AI in vulnerability discovery escalating now?
The partnership arrives against a backdrop of threat analyses warning that AI is compressing the timeline between vulnerability discovery and exploitation.
The Trellix Advanced Research Center's November 2026 CyberThreat Report noted that Anthropic's Claude Mythos Preview, introduced in April 2026, demonstrated an ability to autonomously discover previously unknown vulnerabilities and develop working exploits. Trellix assessed that this signals a fundamental shift in the speed and economics of vulnerability research. The research firm's warning is blunt: the ability to discover vulnerabilities may soon outpace the industry's capacity to validate, prioritize, and remediate them.
ENISA's 2026 State of Cybersecurity in the Union report reached a similar conclusion. The agency found that "frontier AI is having a tremendous impact on vulnerability management, by narrowing response and patch timings," with the time between discovery and weaponization approaching zero. ENISA assessed the overall cyber threat level to the EU as "substantial" — unchanged from 2024 despite heightened geopolitical conflicts.
Read together, these reports explain the logic of Daybreak. If discovery-to-exploitation time is collapsing, defenders need discovery tools that operate at the same speed. OpenAI is supplying those tools to allied governments before adversaries widen the gap.
What remains unclear?
OpenAI's blog post leaves several material questions open. It does not disclose:
- The financial terms of the NATO arrangement
- The specific models made available through Daybreak
- Whether the program includes indemnification or liability provisions for downstream effects of AI-assisted vulnerability research
- How many personnel across NATO and allied agencies have been trained or cleared to use the tools
The liability question deserves attention. AI-assisted vulnerability research produces findings that vendors must act on, and errors at machine speed could carry consequences at machine speed. No published terms address who bears that risk.
What comes next?
NATO's stated intent to study token economics suggests the Alliance will treat Daybreak as input to larger procurement decisions rather than a one-off arrangement. With ENISA's verification data showing a 95% confirmation rate and Ukraine, France, Germany, and Poland already using the models, the program has moved from experiment to operational infrastructure. The open question is whether disclosure practices — on pricing, model versions, and liability — will catch up before the next procurement cycle.
Source: epyhia.com
More from Daniel Okafor
Show full bio
Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.
2,189 articles