OpenAI Brings Frontier Cyber Models to NATO Defenders

Introduction
OpenAI has published an account of how it is putting frontier AI models into the hands of government and national cyber defenders, with NATO and several European agencies among the earliest users. The work runs through a programme called Daybreak, which the company describes as a way to give trusted defenders access to advanced models for authorized security work: reviewing software, investigating suspicious activity, validating vulnerabilities, and developing and testing fixes.
The announcement matters less for the technology itself than for the operational evidence attached to it. Rather than a generic capability claim, OpenAI has pointed to specific engagements with NATO's Communications and Information Agency (NCIA), the EU's ENISA, CERT Polska, and Ukraine's Ministry of Digital Transformation, each with concrete outputs. For security and engineering teams, the interesting question is not whether AI can find bugs, but whether the findings survive human scrutiny and turn into patches that actually stop attacks.
Inside the Alliance: Daybreak and NATO's core networks
According to OpenAI, it works with NATO's NCIA to provide the Alliance's core institutions with access to advanced cybersecurity models, with the stated aim of helping protect Allied networks. The company says NATO has used up to 300 billion raw input and output tokens per month as part of this work.
That figure is a usage measure rather than a security outcome, and it should be read that way. It indicates sustained, high-volume interaction with models across software review and analysis tasks, not a count of vulnerabilities found or incidents prevented. Still, a volume of that scale inside a multinational defence organization suggests the tooling has moved past pilot status and into routine workflows.
OpenAI frames the engagement as more than a licensing arrangement. It says the work helps build cyber expertise across the multinational teams responsible for protecting Allied networks around the clock, and generates lessons that can be shared with defenders across Europe and the wider Alliance. That knowledge-transfer angle is arguably the more durable benefit: model access can be procured quickly, but the tradecraft for using it responsibly inside a security operations centre takes longer to mature.
From exercises to national agencies: what the early results show
The most substantive evidence comes from individual agencies that have gone public with findings.
- ENISA used OpenAI models to identify four vulnerabilities in software used across EU institutions. Fixes were published for all four, meaning the disclosures completed the full cycle from discovery to remediation.
- A November joint report from ENISA and CERT-EU describes a review of 27 software projects using frontier models, including OpenAI models accessed through Daybreak, alongside tools such as the Codex security plug-in. Across the models tested, 95% of the 88 potential issues assessed by software maintainers were confirmed as genuine.
- CERT Polska used OpenAI models to help identify six vulnerabilities in MikroTik RouterOS and coordinated disclosure with the vendor. Researchers directed the work, tested findings on real systems and verified impact, with AI accelerating analysis and the exploration of potential vulnerabilities. MikroTik released fixes, which CERT Polska confirmed prevented the attacks it had observed.
- Ukraine's Ministry of Digital Transformation is using Daybreak to strengthen the cyber defence of civilian infrastructure, aiming to help teams identify software vulnerabilities and develop and test fixes faster amid persistent attacks on essential services.
The CERT Polska case is the most instructive of the four. A confirmed fix that blocks observed attack traffic is a measurable outcome, and it demonstrates the workflow the agencies keep emphasizing: AI proposes, humans verify on live systems, vendors patch, defenders re-test. The MikroTik result also has immediate practical value for network operators running affected RouterOS devices, since vendor fixes are available.
Beyond NATO, OpenAI says it has provided access to its cyber models to defenders in Europe including ENISA and national agencies in France (ANSSI), Germany and Poland.
Human oversight, agentic tooling and what is not yet public
OpenAI also supported the 2026 Crossed Swords exercise, organized by the NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE). The exercise included a new training stream in which participants built, deployed and evaluated AI agents for cyber operations using Daybreak in the technical environment. Participants worked alongside AI agents as operators, with an emphasis on maintaining human oversight.
That framing is deliberate. Agentic security tooling raises obvious questions about scope creep, authorization and accountability, and the exercise structure — build, deploy, evaluate, with humans in the loop — reflects an attempt to rehearse those controls before they are needed in a live environment. OpenAI says it remains committed to NATO's principles for responsible AI adoption.
Several details are not yet public. OpenAI has not disclosed which model versions are involved, the specific CVE identifiers behind the ENISA and CERT Polska findings, or independent benchmarking of false-positive rates beyond the 95% figure drawn from maintainer assessments. Teams evaluating similar tooling should treat vendor-reported accuracy as a starting point, not a procurement decision.
Grivyonx Expert Analysis
The strategic signal here is that AI-assisted vulnerability discovery is being institutionalized inside allied governments rather than trialled at the edges. The 95% confirmation rate across 88 issues is encouraging but also narrow: it reflects findings that maintainers were able to assess, which is a self-selecting sample. The more meaningful metric is time-to-fix and whether patched code survives retesting, and only the CERT Polska result speaks to that directly. Defenders should also note the disclosure pipeline implications. If AI compresses discovery timelines, vendor security teams and coordinators face proportionally more reports per unit time, and triage capacity becomes the bottleneck rather than detection. Organizations that pair model-assisted review with disciplined, human-validated disclosure processes will absorb that pressure; those that automate discovery without matching remediation capacity will simply generate noise.
What defenders should do now
Security and engineering leaders do not need a NATO relationship to act on the patterns in this disclosure.
- Patch the named products. Operators running MikroTik RouterOS should confirm they are on a fixed release and verify that previously observed attack traffic is now blocked, as CERT Polska did.
- Keep a human verification gate. Every model-generated finding should be reproduced on a real system and its impact confirmed before it reaches a vendor or a change ticket.
- Track confirmation rates, not finding counts. Measure how many AI-surfaced issues survive maintainer or internal review, and use that ratio to calibrate how much analyst time the tooling deserves.
- Constrain agent scope. Where agents are used for security operations, define authorized targets, log every action and require sign-off before any state-changing step, mirroring the oversight model used in the Crossed Swords training stream.
- Prepare the remediation path. Faster discovery is only valuable if patching, testing and disclosure coordination scale with it.
Conclusion
OpenAI's account shows allied cyber defenders moving from experimentation to routine use of frontier models, with published fixes from ENISA and CERT Polska as the first hard evidence that the approach produces remediated vulnerabilities rather than just interesting leads. The model is consistent across every case: AI accelerates analysis, humans direct and verify, vendors patch. Organizations building comparable programmes would do well to copy that structure rather than the tooling alone, and teams looking to operationalize AI-assisted security across build and cloud environments can draw on Grivyonx's work in AI automation, DevSecOps and cloud security to wire those controls into existing pipelines.
- #OpenAI
- #NATO
- #cybersecurity
- #AI agents
- #vulnerability disclosure
- #ENISA
- #CERT Polska
- #critical infrastructure



