Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarOpenAI Blog→ original

NATO Uses Up to 300 Billion Tokens a Month in OpenAI's AI Tool for Cyber Defense

OpenAI has opened Daybreak, a program that lets AI read code, hunt for security holes and help verify patches, to NATO and European cyber agencies. NATO runs up to 300 billion tokens a month through it (a token is roughly a syllable of text). In a report by ENISA and CERT-EU covering 27 projects, maintainers confirmed 95% of the 88 problems the AI found. That suggests vulnerability discovery can be handed to AI, while verification stays with humans.

AI-processed from OpenAI Blog; edited by Hamidun News
NATO Uses Up to 300 Billion Tokens a Month in OpenAI's AI Tool for Cyber Defense
Source: OpenAI Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceOpenAI Blog ↗
Editor-in-Chief
Key metric #1With the models' help, ENISA found four vulnerabilities in software used by EU institutions. Fixes have been published for all four.

Imagine your product has hundreds of thousands of lines of code, and nobody to check it for holes: security specialists are scarce, and vulnerabilities are plentiful. Governments face the same problem, except instead of one product they have networks of ministries, hospitals and power grids. OpenAI proposes closing that gap with Daybreak.

It is a program that gives vetted defenders access to advanced models for authorized work: reading code, analyzing suspicious activity, confirming discovered holes and preparing patches. NATO uses it at enormous scale: up to 300 billion tokens a month. A token is roughly a syllable of text that a model reads or writes, and the 300 billion is counted across both incoming and outgoing tokens.

Meanwhile, in a joint ENISA and CERT-EU report, human developers confirmed 95% of the 88 problems the AI found. For practitioners, this is a rare case where AI in security comes with figures from real work rather than from demos.

What NATO Got

The NATO Communications and Information Agency (NCIA) is working with OpenAI so the alliance's key institutions can use advanced cyber models and protect allied networks. Token consumption at this level means AI is built into everyday work, not used for one-off experiments.

NCIA Director General Dylan Brown spoke specifically about money. As usage grows, the agency will study token consumption, use cases and model economics. The goal is to optimize costs, compare competing solutions and make procurement decisions deliberately. A customer at this level is preparing in advance to choose between vendors and count every token.

OpenAI also supported the Crossed Swords 2026 exercises, organized by the NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE). They introduced a new format: participants built, launched and evaluated AI agents for cyber operations on top of Daybreak themselves. Humans worked alongside the agents as operators, with the emphasis on keeping human control.

What the Models Found

Beyond NATO, access to the cyber models went to ENISA (the EU Agency for Cybersecurity), France's ANSSI, and the national cyber services of Germany and Poland. Ukraine's Ministry of Digital Transformation is using Daybreak to protect civilian infrastructure, with the aim of finding vulnerabilities in software faster, developing fixes and verifying them.

The publication lists these concrete results:

  • With the models' help, ENISA found four vulnerabilities in software used by EU institutions. Fixes have been published for all four.
  • The joint November report from ENISA and CERT-EU describes a review of 27 software projects. Teams used frontier models, including OpenAI models via Daybreak, and tools such as the Codex security plugin. Maintainers, the people who look after these projects, assessed 88 potential issues, and 95% of them turned out to be real.
  • CERT Polska used the models to find six vulnerabilities in MikroTik RouterOS, the firmware of network routers, and coordinated disclosure with the vendor. Researchers directed the work themselves, verified findings on real systems and assessed their impact, while the AI sped up analysis and the search through possible holes. MikroTik released patches, and according to CERT Polska, they stopped attacks the team had observed.

The 95% figure matters for a simple reason. The main pain of automated vulnerability discovery is the flood of false alarms, whose triage eats up people's time. When nearly all the problems found are confirmed, AI becomes an assistant rather than a source of noise.

What This Means for Practitioners

First: the "AI proposes, human verifies" model works. In every case described, the final word stays with the researchers and maintainers who test findings on real systems. This is how your own security processes should be built: the model searches for holes and drafts fixes, and a human confirms.

Second: the task is acquiring a market and a price tag. If NATO counts tokens and compares vendors, commercial teams also have reason to measure in advance what one confirmed bug costs.

There are limits too. The source is OpenAI's own publication, so this is a showcase of results, not an independent audit. There are no figures on the share of missed vulnerabilities or on the cost of the work. The 95% figure refers to a verified sample of 88 cases, not to all possible findings. Access went to trusted defenders for authorized work, so it is hard for outsiders to independently check these results for now. But the direction is clear: AI agents for cybersecurity are no longer an experiment and are becoming part of state infrastructure.

Story

Key Questions Answered

What NATO Got?

The NATO Communications and Information Agency (NCIA) is working with OpenAI so the alliance's key institutions can use advanced cyber models and protect allied networks. Token consumption at this level means AI is built into everyday work, not used for one-off experiments. NCIA Director General Dylan Brown spoke specifically about money. As usage grows, the agency will study token consumption, use cases and model economics.

What the Models Found?

Beyond NATO, access to the cyber models went to ENISA (the EU Agency for Cybersecurity), France's ANSSI, and the national cyber services of Germany and Poland. Ukraine's Ministry of Digital Transformation is using Daybreak to protect civilian infrastructure, with the aim of finding vulnerabilities in software faster, developing fixes and verifying them. The publication lists these concrete results: With the models' help, ENISA found four vulnerabilities in software used by EU institutions.

What This Means for Practitioners?

First: the "AI proposes, human verifies" model works. In every case described, the final word stays with the researchers and maintainers who test findings on real systems. This is how your own security processes should be built: the model searches for holes and drafts fixes, and a human confirms. Second: the task is acquiring a market and a price tag. If NATO counts tokens and compares vendors, commercial teams also have reason to measure in advance what one confirmed bug costs. There are limits too.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…