OpenAI said on November 17, 2026, that it is giving AI tools to teams working on NATO cyber defense through the alliance’s technology agency. The work gives NATO’s core institutions models to help check software and test fixes for flaws.
The company works with the NATO Communications and Information Agency, known as NCIA, through its Daybreak program. But reports from European security teams show that finding bugs is only part of the work. Experts still need to check the results.
NATO cyber defense training adds AI agents
The NATO Cooperative Cyber Defence Centre of Excellence said its Crossed Swords exercise used Daybreak for the first time. The exercise ran in Tallinn, Estonia, from November 2 to 6, 2026. A new training stream asked participants to build and assess AI agents, software that can carry out tasks using tools.
According to the center, people worked alongside the agents and kept control of their work. The wider exercise also brought together more than 250 participants from over 40 NATO and non-NATO states.
EU review shows what the results mean
A joint report from the EU Agency for Cybersecurity, ENISA, and the EU institutions’ security service, CERT-EU, describes a separate code-review experiment. It used models from both OpenAI and Anthropic. The report separates possible flaws from those that software owners had checked.
| Measure | Result |
|---|---|
| Codebases reviewed | 27 |
| Potential issues flagged | 220 |
| Findings assessed by code owners | 88 of 220 (40%) |
| Share confirmed among assessed findings | 95% of the 88 reviewed |
Source: ENISA and CERT-EU
The 95% confirmation rate covers the assessed subset, not all 220 potential issues.
The report says the models read source code but lacked details of how the software ran in use. They did not test live apps. The 95% confirmation rate therefore applies only to the reviewed subset; it is not a score for NATO cyber defense or OpenAI alone.
For example, ENISA and CERT-EU said AI could rate a flaw as too severe when it lacked that context. As a result, checking, ranking and fixing the flaws still took expert time.
Polish researchers tested flaws on real routers
Similarly, CERT Polska, Poland’s national incident-response team, shows how those checks work. It found six flaws in MikroTik RouterOS, the software used on the company’s routers. Researcher Sławomir Rozbicki used OpenAI models in a lab with the team in control.
According to CERT Polska, each possible flaw needed tests on a real RouterOS system. The team then repeated checks on a clean machine and assessed the impact. AI sped up analysis and the search for weaknesses. But setting up the lab, ruling out false leads and recording the results remained major parts of the work.
The team also confirmed attacks exploiting two of the flaws and said patches blocked the observed attacks. Its advice also calls for checks for unknown accounts or other changes made without permission.
OpenAI says it will continue working with governments and security agencies. In a statement published by the company, NCIA General Manager Dylan Browne said the agency would study usage and costs to guide future purchases.
Sources: OpenAI, NATO CCDCOE, ENISA and CERT-EU, CERT Polska
