OpenAI is opening its Daybreak cybersecurity system to NATO and government teams in France, Germany and Poland. The tool uses AI models to inspect software, identify vulnerabilities and help defenders test patches under human oversight.

European agencies have reported early operational results. ENISA and CERT-EU examined 27 software projects and said 84 of 88 issues reported by the system—95%—were genuine bugs. Poland’s CERT Polska used the technology to find six vulnerabilities in MikroTik routers that attackers were already exploiting. Ukraine is also applying Daybreak to civilian infrastructure.

The NATO Communications and Information Agency says it already uses the underlying models extensively. During the Crossed Swords 2026 exercise, participants deployed AI agents for cyber operations under human supervision for the first time. Access now expands from model use to the packaged defensive workflow.

The reported validation rate is promising but comes from selected agency projects rather than a broad independent benchmark. Security teams still need to reproduce findings, prioritize severity and review patches before deployment. Models capable of finding flaws can also be misused, so access controls and monitoring matter as much as raw detection. Daybreak’s near-term value is faster code auditing for trusted defenders, with humans retaining responsibility for deciding which findings are real and how to fix them safely.