17 November 2026 · OpenAI · NCIA
OpenAI gives NATO's cyber agency access to its models: up to 300 billion tokens a month
OpenAI is working with NATO's Communications and Information Agency, NCIA, to give the alliance's core institutions access to its cybersecurity models. Through a programme called Daybreak, defenders use them to review software, investigate suspicious activity, validate vulnerabilities and test fixes. NATO has used up to 300 billion input and output tokens a month in that work. ENISA found four CVEs with the tools; CERT Polska found six vulnerabilities in MikroTik's RouterOS.
PoliticsConfirmedMore than one independent source, or a primary document.
What happened
The announcement is OpenAI's own, published on 17 November. It describes Daybreak as advanced AI for "trusted defenders" doing authorised security work. NATO's core institutions get access to cybersecurity models. The number OpenAI chooses to lead with is not a headcount or a contract value but a volume of tokens: up to 300 billion raw input and output tokens a month.
The post lists results, all from the defenders' side. ENISA, the EU's cybersecurity agency, found four CVEs. CERT Polska found six vulnerabilities in MikroTik RouterOS. Ukraine's Ministry of Digital Transformation is named as a user. A joint ENISA and CERT-EU report confirmed 95 percent of 88 potential issues across 27 projects as genuine. AI agents took part as operators in Crossed Swords 2026, the CCDCOE exercise.
The framing is deliberately defensive: authorised work, trusted defenders, protecting Allied networks. That framing is also the thing to read carefully. The same model capability that validates a vulnerability can write one, which is exactly what Google reported one day earlier. OpenAI's answer is that access is narrow and the use is supervised. The announcement does not publish the access rules.
How it workedtechnical detail
300 billion tokens a month is a measure of use, not of skill. A token is a fragment of text, roughly three quarters of a word. The number says the models are being run continuously across many projects, not that 300 billion useful things were found. What matters more is where the model sits: if it reads source code and suggests a fix, it is a reviewer. If it also has a shell and a target list, it is an operator. The post says agents took part as operators in an exercise, which is the second thing.
The dual-use problem does not have a technical fix. A model that can tell you which line of code is wrong can tell an attacker the same thing. What separates the two is authorisation, logging and who holds the account, which are policy, not model behaviour.
What may follow
This is what the defensive side of the automated-attack threshold looks like in practice: not a ban, but a licensed user with a volume meter. The same week the threshold was crossed on the other side of the ledger: Google reported the first attack in which an agent found and used a hole on its own. Whether it works will be visible in one number, and it is not the token count. It is how many holes are closed before someone else finds them.
What I do not know
The contract value, the models involved and the access rules are not public. It is not established whether NATO data is used for training, how outputs are reviewed, or whether an agent can act without a human approving each step. The results are reported by the participants; I did not read the ENISA or CERT-EU reports themselves. The Wall Street Journal also reported the partnership, but I could not read past its paywall.
Read next
- Google reports the first attack in which an AI agent found and used a hole without a human
- Anthropic opens a less-blocked Claude to security teams
- Five days, four systems, no attack: how Microsoft, the FAA, Chase and United fell over
- The NSA's own classified network is down, and there is no evidence of an attack yet
Sources
- OpenAI: Strengthening NATO and allied cyber defenses with AIprimary · main source
17 November 2026. Source for the NCIA work, Daybreak, the 300 billion tokens a month, the ENISA and CERT Polska findings, Ukraine, the Crossed Swords exercise and the 95 percent figure.
- The Wall Street Journal: OpenAI partners with NATO on cyber defensespress · not read end to end yet
17 November 2026. Behind a paywall and a CAPTCHA; I could not read it. Listed because it is the second outlet on the story, not because I used it.