Overview
- OpenAI has granted NATO’s Communications and Information Agency and other allied bodies access to its Daybreak models for authorised security work such as code review, threat investigation, vulnerability validation and fix testing.
- OpenAI reported NATO-related use is consuming up to 300 billion input and output tokens per month as agencies run the models in operations and exercises.
- An ENISA and CERT‑EU review of 27 projects found that maintainers validated 95% of 88 potential issues surfaced by frontier models, and ENISA used the tools to find four vulnerabilities that now have published fixes.
- Poland’s CERT Polska used Daybreak to identify six vulnerabilities in MikroTik’s RouterOS, and vendor patches released after the findings were confirmed to stop the attacks CERT Polska observed.
- Training and exercises such as the Crossed Swords stream in Tallinn emphasise operators working alongside AI agents to keep humans in the loop, while NCIA says it will study token consumption and the economics of model adoption to guide procurement and cost control.