OpenAI Expands Daybreak Across NATO and Allied Agencies | Zeniteq
Loading
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
OpenAI Expands Daybreak Across NATO and Allied Agencies
NATO’s reported AI usage shows the scale of Daybreak’s deployment, while European vulnerability reviews offer a separate measure of its defensive value.
OpenAI has expanded its Daybreak cybersecurity platform across NATO and allied agencies, with NATO institutions consuming up to 300 billion raw input and output tokens monthly. While this high volume indicates significant adoption, it measures processing capacity rather than security efficacy. More concrete evidence comes from European agencies, where researchers used these models to identify vulnerabilities, such as six flaws in MikroTik RouterOS. Crucially, defenders verified that vendor patches successfully blocked these specific threats. This distinction highlights that while token usage tracks procurement scale, actual defensive value relies on human-led validation and fix verification.
NATO institutions have used OpenAI’s cybersecurity models through Daybreak at volumes of up to 300 billion raw input and output tokens per month, according to OpenAI. In its November 17, 2026 announcement, the company described access through NATO’s Communications and Information Agency, or NCIA, alongside deployments involving European and Ukrainian cyber defenders.
The announcement also reports concrete security outcomes: software maintainers confirmed 95% of 88 assessed potential issues in an ENISA and CERT-EU review, while CERT Polska helped identify six MikroTik RouterOS vulnerabilities and verified that vendor fixes blocked attacks demonstrated during its work. Those findings describe different projects, with different participants and validation methods. They should be read separately from NATO’s usage figure. (OpenAI)
That distinction is the most useful way to assess this expansion. Token consumption measures how much model processing an institution uses. Maintainer confirmation measures whether submitted findings describe genuine problems. Testing a fix adds another layer: evidence that a specific attack no longer works under the conditions researchers checked.
NATO’s 300 Billion Tokens Measure Usage, Not Results
OpenAI says NCIA provides NATO’s core institutions with access to advanced cybersecurity models through Daybreak to help protect Allied networks. The company describes authorized work spanning software review, investigation of suspicious activity, vulnerability validation, and development and testing of fixes. (OpenAI)
Tokens are the units a model processes and generates, including pieces of text and code. The reported 300 billion figure combines raw input and output tokens. It is therefore neither a count of security tasks nor a count of vulnerabilities found.
Three details matter when interpreting that number:
“Up to” describes reported usage at the high end. It does not give a monthly average or establish a contractual usage limit.
Input and output are combined. The figure does not reveal how much material defenders supplied versus how much the models generated.
No task breakdown accompanies it. Readers cannot divide the total into code review, investigations, training, or patch development.
OpenAI’s announcement does not specify a NATO deployment start date, model-version breakdown, pricing, or cost per completed task. also notes the missing start date.
The volume is consequential for procurement because it suggests substantial model consumption across institutional work. But converting it into a bill would require pricing and workload details that the announcement does not provide. A long software review and a short triage request can consume very different amounts of input, and repeated analysis can increase usage without producing an additional confirmed finding.
My view is that the token figure belongs in a capacity and cost discussion. It is useful evidence of adoption, but it cannot answer whether the models improved security or saved defenders time.
Crossed Swords Put Operators Beside AI Agents
OpenAI also supported the 2026 Crossed Swords exercise, organized by the NATO Cooperative Cyber Defence Centre of Excellence. A new training stream had participants build, deploy, and evaluate AI agents for cyber operations using Daybreak in the technical environment. OpenAI says participants worked alongside those agents as operators, with an emphasis on human oversight. (OpenAI)
The training description matters because using an agent involves more than asking a model a single question. Participants were learning to deploy and assess systems that perform cyber tasks, while retaining an operator role. The announcement does not detail the agents’ permissions, execution environments, or approval gates, so it would be premature to describe the exercise as unrestricted autonomous cyber operations.
A controlled exercise can help teams practice supervising model behavior before relying on it in operational workflows. That is a practical inference from the training format, not a reported performance result. OpenAI provides no exercise score, comparison with unaided teams, or measurement of time saved.
NCIA General Manager Dr Dylan Browne also framed adoption as an ongoing procurement exercise. In the announcement, he said greater understanding of usage economics would support decision-making:
“This insight will help us optimise costs, evaluate competing solutions and make informed procurement decisions.”
The statement leaves an important question open. Even at NATO’s reported volume, its leadership is still describing the need to understand consumption patterns and compare alternatives.
What the European Vulnerability Results Actually Show
The strongest outcome evidence in the announcement comes from human assessment and fix verification. OpenAI describes three related but distinct pieces of work:
Work described by OpenAI
Scope and reported result
What the result supports
ENISA software investigation
Four vulnerabilities found in software used across EU institutions; fixes published for all four
A high confirmation rate among the issues maintainers assessed
CERT Polska RouterOS research
Six vulnerabilities identified with model assistance; vendor fixes tested against observed attacks
Findings were investigated on real systems, and the relevant fixes passed the reported checks
These are OpenAI’s published accounts of agency work. The opened evidence includes its announcement and coverage from The Decoder and Quartz, which largely relay the same account. That reporting does not independently reproduce the agencies’ tests or validate OpenAI’s figures.
The 95% Rate Covers Assessed Issues Across Multiple Models
OpenAI says the November joint ENISA and CERT-EU report reviewed 27 software projects using frontier models, including OpenAI models accessed through Daybreak. Tools included the Codex security plug-in. Across the models tested, software maintainers confirmed 95% of 88 assessed potential issues as genuine. (OpenAI)
The 95% figure is not a Daybreak-only accuracy score. It pools findings across the models tested, and the denominator is the 88 potential issues maintainers assessed.
That scope limits several tempting interpretations. The figure does not say the systems found 95% of all vulnerabilities present in the projects. It also does not reveal how many candidate findings were generated before the assessed set was selected, or how much analyst effort went into checking them.
Maintainer confirmation is still a useful measure. It moves a finding beyond a model’s unsupported assertion and records that the software’s maintainers accepted it as genuine. But it measures the validity of assessed reports, not the completeness of the search.
The four ENISA vulnerabilities with published fixes are a separate result. The announcement does not say every confirmed issue in the broader 27-project review had been patched. Keeping those counts separate prevents a validation statistic from turning into an unsupported claim about remediation.
MikroTik Fixes Add a Practical Security Outcome
CERT Polska’s RouterOS work provides a different kind of evidence. According to OpenAI, researchers used its models to help identify six vulnerabilities, directed the investigation, tested findings on real systems, and verified their impact. They coordinated disclosure with MikroTik, which released fixes. CERT Polska then confirmed that those fixes prevented the attacks it had observed. (OpenAI)
The sequence is more informative than a vulnerability count alone:
Models assisted analysis and exploration of possible weaknesses.
Researchers tested the findings and checked their impact.
The team coordinated disclosure with the vendor.
MikroTik issued fixes.
Researchers checked whether the fixes blocked the relevant attacks.
For a network operator, that final check is especially valuable. A plausible finding can create investigation work; a validated fix gives defenders a concrete remediation path.
OpenAI’s summary does not provide the affected RouterOS releases or the versions containing the fixes. Operators therefore need the relevant MikroTik and CERT Polska advisories to determine applicability before changing production systems. The account also should not be read as proof that all six vulnerabilities were being exploited in the wild: attacks observed or demonstrated during research are a narrower claim.
I would give this fix-validation evidence more weight than raw usage when assessing practical defensive value. It connects model-assisted research to a change in software behavior, while keeping researchers responsible for testing the result.
Allied Access Extends Beyond NATO
OpenAI says it has also provided access to cyber models to ENISA and national cyber agencies in France, Germany, and Poland. Its announcement identifies France’s ANSSI and describes Ukraine’s Ministry of Digital Transformation as using Daybreak to strengthen the cyber defense of civilian infrastructure.
The Ukrainian initiative is intended to help teams identify software vulnerabilities and develop and test fixes more quickly while protecting essential services amid persistent cyberattacks. (OpenAI)
The public account is uneven in its level of detail. NATO has a usage figure, ENISA and CERT-EU have assessed findings, and CERT Polska has a documented remediation sequence as summarized by OpenAI. The announcement provides no comparable quantitative outcome for Ukraine, France, or Germany.
Access, deployment volume, and validated results are separate milestones. Grouping every participating agency under one success rate would erase those differences.
Buyers Need Usage and Outcome Measures Together
The announcement supports a practical procurement lesson: track consumption and verified outcomes side by side, within the same workflow.
NATO’s token count and the ENISA/CERT-EU confirmation rate cannot be combined into a cost-per-vulnerability calculation. They come from different work, and the necessary pricing, attribution, and labor data are absent.
A more useful assessment would record:
Model and tool versions, input/output usage, and cost by task.
Candidate findings, findings submitted for assessment, and findings confirmed.
Analyst time spent investigating valid and invalid reports.
Fix status and the tests used to verify remediation.
A comparison with the team’s existing workflow, where feasible.
Those measures would help distinguish a model that generates many convincing reports from one that helps defenders reach reliable fixes efficiently. They would also make competing systems easier to compare on equivalent work.
Daybreak’s expansion brings frontier models into consequential national-security settings. The published evidence is most persuasive where people checked findings and tested fixes. NATO’s usage shows scale; the next procurement question is how consistently that scale translates into verified defensive outcomes.
Frequently Asked Questions
3 questions
1
Is NATO using Daybreak autonomously?
OpenAI describes human oversight during Crossed Swords training. Its announcement does not detail permissions or approval controls for NATO’s wider deployment.
2
Is the 95% Confirmation Rate Specific to OpenAI?
No. OpenAI says it covers assessed issues across multiple frontier models in the ENISA/CERT-EU review, including models accessed through Daybreak.
3
Which MikroTik RouterOS Versions Need Updating?
The opened announcement does not list affected or fixed versions. Operators should consult the relevant MikroTik and CERT Polska advisories before selecting an update.