The programme itself is not new. OpenAI launched Daybreak earlier in 2026. The change this week is scope. What started as a vetted pilot now has a permanent institutional customer in NATO and a set of named European agencies that have published early results.
What the defenders actually get
Daybreak is not a chatbot with a security prompt. It is a framework for giving authorised defenders access to frontier models inside their own workflows: auditing source code, inspecting suspicious network behaviour, validating whether a reported vulnerability is exploitable, and testing patches before they ship.
Access is split into two tiers. Daybreak Blue covers defensive work such as code review and incident triage. Daybreak Red covers penetration testing and red teaming, which is why it sits behind stricter vetting. Both tiers require human-in-the-loop oversight, according to the programme description. That is not a marketing detail. It is the thing that makes the model output usable at all in a security context, because an analyst still has to sign off before anything becomes a ticket, let alone a patch.
ENISA found 88 issues across 27 projects
The most concrete public result comes from ENISA and CERT-EU. The agencies evaluated 27 software projects using the models. Maintainers then assessed 88 potential issues the tools had flagged and confirmed that 95% were real. That work led to patches for 4 major vulnerabilities in software used by EU institutions.
Read that 95% carefully. A security tool that is wrong one time in twenty is still a tool that costs a maintainer time, and false positives are the usual reason scanner output gets ignored. Confirmations came from the maintainers themselves, not from the model, which is the part of this workflow worth copying: the model proposes, a human with commit rights disposes.
The Polish CERT results are narrower but more specific. It used OpenAI models to find 6 vulnerabilities in MikroTik RouterOS, and vendor fixes for them were verified and released. RouterOS runs on hardware in a lot of European small business networks, which is exactly the kind of software that rarely gets a funded audit.
300 billion tokens a month, priced out
NATO's contract terms are not public. The volume is: NCIA workflows consume up to 300 billion input and output tokens per month, according to the disclosed figures.
That number is worth converting, with the obvious caveat that a subsidised programme probably does not pay list rates. Assume a 9:1 split between input and output, which is plausible for code auditing where large files go in and short findings come out, so 270 billion input tokens and 30 billion output tokens per month:
| Illustrative price band | List price (input / output per 1M) | Rough monthly cost at 300B tokens |
|---|---|---|
| Low band | $2.00 / $10.00 | ~$0.84M (≈ €0.78M) |
| Mid band | $8.00 / $40.00 | ~$3.36M (≈ €3.1M) |
| High band | $20.00 / $100.00 | ~$8.40M (≈ €7.8M) |
The euro figures assume about $1.08 per euro and are there for scale, not as a quote. Even at the top of that range, roughly $100 million a year, this is a list-price scenario for model usage, not a programme budget. That arithmetic says something about where the money in defender AI actually goes: people, integration, training and support, not inference. Ukraine uses Daybreak to protect civilian infrastructure.
Your company cannot sign up for this
Daybreak access runs through agencies and vetted partners. There is no self-serve form, and the commercial terms for private defenders have not been published. A European security team that wants the same capability today has two realistic routes.
The first is the standard OpenAI API. The frontier models are available to EU customers at list rates, and a team can build its own review pipeline on top of a model priced at the low illustrative band above: $2 per million input tokens and $10 per million output tokens. The second route matters more for anyone who cannot send source code to a US endpoint: Mistral Large 4 ships open weights under the Mistral Ecosystem Open-Weight License, at $0.70 per million input tokens and $2.10 per million output on the API. That is roughly a third of the input price and a fifth of the output price of the low band, and the weights can run inside an EU datacentre you control.
Self-hosting a flagship-class model is a multi-GPU job, not something that fits next to your IDE, and we track what actually fits on consumer hardware in AI Arena. For a CERT or an internal security team handling unreleased code, though, jurisdiction is often the deciding factor, not tokens per second.
The GDPR and AI Act questions nobody answered this week
Two things are missing from the published material. It does not say where inference runs for the European agencies, and it does not describe the data-processing terms. For a public body feeding source code and network telemetry into a model, those are the questions that decide whether the deployment survives legal review.
The AI Act side is clearer in one respect. Chapter V obligations for general-purpose model providers sit with OpenAI, not with the agency buying access, and the code of practice applies regardless of customer. What the deploying agencies own is procuring a system with documentation, keeping logs, and being able to show a human made the call before a patch shipped. Daybreak's stated human-in-the-loop requirement matches that expectation, though the audits that would prove it have not been published.
The public disclosure landed alongside the NATO Edge 26 programme, and NCIA has not released per-agency token figures. During NATO CCDCOE's Crossed Swords 2026 exercise, participants deployed AI agents for cyber operations under human oversight for the first time. Those exercise results have not been published either.
Does Daybreak replace penetration testers or code reviewers?
The programme requires human-in-the-loop oversight, and the ENISA figures show why: maintainers had to assess each flagged issue themselves, and about 5% of the 88 did not hold up. The tool shortens the search, it does not sign off the finding.
Can a private European company get Daybreak Blue or Red?
No public route exists. Every disclosed participant is a government agency, NATO body or vetted partner, and OpenAI has not published commercial access terms for private defenders. The models themselves remain available through the standard API.
Is the token volume comparable to normal API usage?
Not remotely. 300 billion tokens per month is roughly the output of a large enterprise running continuous code analysis across thousands of repositories. At the low illustrative band above, it would cost a commercial customer close to $0.84 million per month.