OpenAI says NATO’s core institutions are now defending allied networks with its Daybreak cybersecurity models — burning through up to 300 billion tokens a month — and that cyber agencies in the EU, France, Germany, Poland and Ukraine are using the same programme to find real vulnerabilities. Why it matters: a frontier AI lab is now named, on the record, as part of NATO’s working cyber-defence toolkit.
November 17, 2026 — İzmir / Brussels
What happened
In a Global Affairs post published Tuesday, OpenAI said it works with the NATO Communications and Information Agency (NCIA) to give NATO’s core institutions access to its advanced cybersecurity models through Daybreak — the programme that hands OpenAI’s most capable cyber models to vetted defenders for authorised work: reviewing software, investigating suspicious activity, validating vulnerabilities and building and testing fixes. OpenAI put NATO’s usage at up to 300 billion raw input and output tokens per month. Neither OpenAI nor the NCIA disclosed a price or contract value, and NCIA general manager Dylan Browne signalled the alliance will study token use and costs to evaluate competing solutions before committing further. The announcement landed as NATO Edge 26, the NCIA’s annual industry conference, opened in İzmir, Türkiye.
The European results OpenAI is pointing to
OpenAI also listed results from agencies already on Daybreak. The EU’s cybersecurity agency ENISA used the models to find four vulnerabilities in software used across EU institutions, with fixes published for all four. A joint ENISA and CERT-EU review of 27 software projects, published November 12, gives the fuller picture:
- The models flagged 220 potential issues across the 27 codebases.
- Code owners had assessed 88 findings by publication and confirmed 95% of them as real; four new CVE IDs were assigned.
- The work used about 611 million tokens at an estimated €6,649 — although ENISA received the models free under the companies’ access programmes.
- The catch: the models tended to overrate the severity of their findings, and checking them took expert time ENISA says is scarce.
Poland’s CERT Polska used the models to help find six flaws in MikroTik’s RouterOS router software. Ukraine’s Ministry of Digital Transformation — on Daybreak since September — uses it to protect civilian infrastructure. Access also extends to France’s ANSSI and national cyber agencies in Germany and Poland.
AI agents at NATO’s big cyber exercise
OpenAI also supported Crossed Swords 2026, the exercise run by NATO’s Cooperative Cyber Defence Centre of Excellence in Tallinn from November 2 to 6. More than 250 participants from over 40 countries took part, and this year the exercise added a training stream in which participants designed, deployed and evaluated AI agents for cyber operations — with human operators working alongside the agents, and Daybreak in the technical environment for the first time.
Why it matters
Context makes this more than a customer logo. On Monday, Google’s threat team reported attackers using an AI framework that turned security patches into working exploits with no human in the loop — the defensive case for giving vetted teams the strongest cyber models is getting easier to make, and NATO is now making it publicly. The honest caveats are in the EU’s own report: the models find real bugs fast, they also inflate severity, and scarce human experts still have to check everything. And the dependency question is now explicit — allied governments are running a U.S. company’s models over their own networks, with the alliance openly reserving the right to shop around. Everything AI, tested: this time, tested in uniform.
FAQ
What is OpenAI Daybreak?
Daybreak is OpenAI’s programme for giving vetted, authorised defenders access to its most capable cybersecurity models. Tested Intelligence covered its Red Team and Defense Factory products when they launched on November 12.
How much is NATO paying OpenAI?
Neither side disclosed a price or contract value. OpenAI cited usage of up to 300 billion tokens a month, and the NCIA says it will study token use and costs to evaluate competing solutions.
Did the models actually find real vulnerabilities?
Yes, per the agencies themselves: ENISA found four vulnerabilities in EU-institution software (all fixed), its joint review with CERT-EU confirmed 95% of 88 assessed findings as real, and CERT Polska credited the models with helping find six MikroTik RouterOS flaws.
Sources: OpenAI, ENISA and CERT-EU, NATO CCDCOE (reporting via Mad Robot and Times of AI)

