An upper-end figure, presented as scale
Three hundred billion is the number this announcement wants quoted, so it is worth reading the words around it. According to Help Net Security's account of OpenAI's announcement, NATO has used up to 300 billion raw input and output tokens a month. "Up to" gives an upper-end monthly figure; the reporting does not identify the period it covers, an average, or a typical month. The tokens are raw input and output combined, which means the code, logs and context fed into the models count alongside everything they produced. A code-review workflow that re-reads a large repository on every pass generates tokens at a rate that says a lot about the size of the repository and very little, on its own, about what was found in it.
The total establishes reported consumption. It does not, on its own, establish deployment maturity, waste, or how the vendor is paid — none of the commercial terms are disclosed.
The buyer is still studying the economics
The most useful sentence in the announcement comes from the customer. NCIA general manager Dr Dylan Browne said that as adoption scales, the agency will build a deeper understanding of token consumption, usage patterns and the underlying economics of the models, so it can optimize costs, evaluate competing solutions and make informed procurement decisions.
Read carefully: Browne describes further analysis of usage and economics as adoption scales. The reporting does not disclose what assessments NCIA has already done, or whether they are complete, and no one quoted in this announcement, on either side, is yet claiming the program is cost-effective. That question is left open by the customer's own words.
What can actually be counted
The same announcement contains results that do not need a token denominator, reported separately from NATO's usage. ENISA found four vulnerabilities in software used across EU institutions, and fixes are published for all four. Separately, CERT Polska's researchers used the models to help find six vulnerabilities in MikroTik RouterOS, directed the work themselves, tested the findings on real systems, and confirmed that MikroTik's fixes stop the attacks they had observed. Neither result is presented as an outcome of NATO's token expenditure, and together they do not exhaust what Daybreak's users may have done; they are the countable examples the announcement offers.
A third figure sits between the two kinds. ENISA and CERT-EU reviewed 27 software projects, and maintainers confirmed 95% of the 88 assessed potential issues in that multi-model review as genuine. That confirmation rate should be quoted with its conditions attached: the review used several frontier models, not OpenAI's alone, and the selection process behind the 88 assessed issues does not establish performance across all model output.
An exercise, with the humans named
Daybreak also appeared at Crossed Swords 2026, the CCDCOE exercise where participants built, deployed and evaluated AI agents for cyber operations inside the Daybreak environment. The reporting of the exercise stresses that human operators worked alongside the agents and retained oversight. That emphasis matches the pattern in the CERT Polska result above, where a named human team directed the work and checked it on real systems.
What the announcement does not contain is any account of NATO's own usage producing a comparable list. NATO's section of the story is access, tasks and token volume. The findings belong to ENISA, CERT-EU and CERT Polska. That may simply reflect what an alliance agency can say publicly about its own networks. It still leaves the largest number in the piece attached to the least specific outcome.
Sources
- [1] OpenAI, "Strengthening NATO and allied cyber defenses with AI", Global Affairs post, November 17, 2026 (vendor announcement; NOT read directly by ALFW — read via the two reports below; stored without a URL because the original's address was not confirmed)
- [2] Help Net Security, "NATO's AI cyber defense work hits up to 300 billion OpenAI tokens a month", November 17, 2026 (read first-hand by ALFW)Read source
- [3] Defence Industry EU, "NATO uses OpenAI models to protect alliance networks as cyber defenders train with AI agents", November 17, 2026 (read first-hand by the drafting desk)Read source