Field noteNov 17, 20265 min read1 source

Strengthening NATO and allied cyber defenses with AI

In one minute

OpenAI says NATO’s Communications and Information Agency has used its cybersecurity models through Daybreak, reaching up to 300 billion input and output tokens per month. The announcement also describes deployments among allied agencies where human defenders direct, test, and validate model-assisted security work.

What changed: OpenAI says NATO’s core institutions use its cybersecurity models through NCIA, with usage reaching up to 300 billion tokens per month, and describes AI-agent training at Crossed Swords.

What happened

OpenAI’s November 17, 2026 announcement describes work with NATO’s Communications and Information Agency (NCIA) to give NATO’s core institutions access to advanced cybersecurity models through Daybreak. OpenAI describes authorized work ranging from software review and investigation of suspicious activity to vulnerability validation, development of fixes, and testing of those fixes. The company says NATO has used up to 300 billion raw input and output tokens per month as part of the work.

The announcement describes a training stream at the 2026 Crossed Swords exercise, organized by the NATO Cooperative Cyber Defence Centre of Excellence. Participants built, deployed, and evaluated AI agents for cyber operations in a technical environment using Daybreak. OpenAI says the training positioned participants alongside AI agents as operators, with an emphasis on maintaining human oversight. It presents the exercise as a way to develop cyber expertise across multinational teams protecting Allied networks.

Beyond NATO, OpenAI says it has already given access to cyber models to European defenders, including the EU Agency for Cybersecurity (ENISA) and national cyber agencies in France, Germany, and Poland. Ukraine’s Ministry of Digital Transformation is using Daybreak to identify software vulnerabilities in civilian infrastructure and to develop and test fixes more quickly, according to OpenAI.

OpenAI says ENISA identified four vulnerabilities in software used across EU institutions with its models, and that fixes were published for all four. It also cites a November joint report by ENISA and CERT-EU that reviewed 27 software projects using frontier models, including OpenAI models accessed through Daybreak and tools including the Codex security plug-in. Across the models tested, maintainers confirmed 95% of 88 potential issues as genuine. Separately, OpenAI says CERT Polska used its models to help identify six vulnerabilities in MikroTik RouterOS. Researchers directed the work, tested findings on real systems, and verified their impact; OpenAI says MikroTik released fixes that CERT Polska confirmed prevented the attacks it had observed.

What the archive adds

New Runtime’s earlier coverage described Daybreak as a way of placing advanced cyber capability with trusted defenders, with access boundaries central to the approach. This is a new fact on that known story: OpenAI now reports institutional deployment, a usage figure, and examples from NATO and allied agencies.

Trend signal

This supports AI-native organizations move toward review and orchestration. OpenAI’s examples place models inside a division of labor in which people set direction, investigate findings on real systems, validate impact, and confirm fixes. The announcement does not establish that the workflow improves outcomes across all institutions, but it does describe review as part of the operating model rather than an afterthought.

Why now

OpenAI frames the announcement around expanding use by NATO and allied agencies. It pairs reported NATO usage with a November ENISA-CERT-EU report and the 2026 Crossed Swords exercise, making this an adoption and training update rather than a newly launched public product.

Why it matters

For teams building with AI, the useful signal is the division of responsibility, not the headline token count. The examples describe models helping teams inspect software, explore possible vulnerabilities, investigate suspicious activity, and propose or test fixes. They also describe specialists directing research, testing findings against real systems, and verifying impact before acting.

That distinction matters for any organization considering agents in high-consequence workflows. A model-generated lead is not equivalent to a confirmed vulnerability, and an apparent fix is not equivalent to a deployed remediation. The reported NATO scale also raises a procurement question: large token consumption indicates use, but not cost-effectiveness. NCIA’s general manager says the organization is seeking a deeper understanding of consumption, usage patterns, and model economics to optimize costs and evaluate competing solutions.

Limits and uncertainty

These are OpenAI’s descriptions of deployments and results, rather than an independent assessment of their overall security impact. The company does not define the measurement window behind NATO’s peak monthly token figure, allocate that usage by task, or report cost, time saved, detection rates, or comparison with prior security workflows. It does not identify model versions, explain access controls or agent permissions, or specify the review process used across the deployments.

The cited vulnerability examples establish that findings were assessed and fixes published, but they do not isolate what the models contributed relative to researchers and existing tools. The ENISA-CERT-EU figure covers models tested across a set of projects, rather than a performance measure for Daybreak alone.

Skill to build now

For any agent-assisted security workflow, define a confirmation gate before deployment: require an authorized reviewer to reproduce the finding on an approved environment, record the impact, and verify the proposed remediation. Use agents to expand investigation and generate leads; use named human owners to decide whether a lead becomes a ticket, a disclosure, or a production change.

Reader takeaway

OpenAI’s announcement is evidence of government cyber teams putting frontier models into real security workflows, but not evidence that security can be delegated to a model. The reusable operational lesson is to design around authorization, human verification, and confirmed remediation—not model output alone.

Open question

Will NATO or an allied agency publish a reproducible account of the workflow, model access controls, and economics behind this reported usage?

Evidence / sources

  1. 1openai.com/index/strengthening-nato-allied-cyber-defences-ai
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicModelsExplore the models topic hub.
  2. 02archiveField NotesOpen the latest editorial analysis.
  3. 03sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract