OpenAI Says NATO Uses Its Cybersecurity Models Through Daybreak Program
NATO institutions get OpenAI cyber tools through its Daybreak program
OpenAI said on Nov. 17 that it is working with NATO's Communications and Information Agency (NCIA) to give the alliance's core institutions access to its advanced cybersecurity models. The access runs through OpenAI's Daybreak program, a system that grants approved users special permission to use the company's AI for security work.
The company said the tools support authorized tasks such as finding software weaknesses, confirming vulnerabilities, and building and testing fixes. According to OpenAI, NATO has already used up to 300 billion raw input and output tokens per month. A token is a small unit of text that AI models process.
The main points
- OpenAI announced the NATO work on Nov. 17 through its Daybreak trusted-access program.
- NATO reportedly used up to 300 billion raw input and output tokens per month.
- The announcement did not name the specific models NATO uses, nor a start date or financial terms.
- NATO's cyber defense center used Daybreak for the first time in its Crossed Swords exercise, held Nov. 2–6 in Tallinn.
- Other agencies with access include ENISA, France's ANSSI, cyber agencies in Germany and Poland, and Ukraine's Ministry of Digital Transformation.
What OpenAI's announcement says
The announcement quoted NCIA General Manager Dylan Browne as saying that understanding usage and costs would help the agency "evaluate competing solutions and make informed procurement decisions."
Separately, the NATO Cooperative Cyber Defence Centre of Excellence confirmed that its Crossed Swords exercise in Tallinn used Daybreak in its technical environment for the first time. Participants built and evaluated AI agents for cyber operations. The organizer stressed that humans kept oversight of the work.
What Daybreak access allows
OpenAI's Daybreak access guide describes two approval-based tiers. The practical difference from standard access is fewer refusals on approved security tasks. At the higher tier, users also get cyber-specialized models. The NATO announcement does not specify which tier NATO uses.
OpenAI says it reviews each request before enabling access, and approval is not automatic. Daybreak is limited to systems users own, operate, or are explicitly authorized to test or analyze. Users cannot share access with outside parties. OpenAI's usage policies also ban malicious cyber activity and unsolicited safety testing, and require review and approval for national-security or intelligence use.
How other cyber agencies use the models
OpenAI also named ENISA, France's ANSSI, and national cyber agencies in Germany and Poland as recipients of access. It did not identify the German agency. The company said Ukraine's Ministry of Digital Transformation uses Daybreak to defend civilian infrastructure.
A November report by ENISA and CERT-EU confirms access to OpenAI's GPT-5.5 and GPT-5.6 families and Astra through Daybreak, with tools including Codex and its security plug-in. The report noted that its 27-codebase experiment also used Anthropic models. It stressed that the reviews tested source code, not live applications, and that models could misjudge severity without deployment context.
CERT Polska reported that its researchers used OpenAI models to help discover six MikroTik RouterOS vulnerabilities. The Polish team said the models sped up analysis, but researchers still had to test every hypothesis on real systems, rule out false conclusions, and assess the impact.
What is confirmed and what is not
The NATO partnership, the token usage figure, and the Crossed Swords exercise use come from OpenAI's announcement and the NATO cyber defense center's own confirmation. The details of Daybreak's rules come from OpenAI's access guide and usage policies.
Several details remain unknown. OpenAI did not say which models NATO uses, which Daybreak tier it holds, when the work started, or what it costs. It also did not identify the German agency that received access.
Why this matters
The announcement shows that a major military alliance is using commercial AI models for defensive cybersecurity work, and that the usage is already at a large scale. It also shows how AI companies are handling government and security work through controlled access programs with approval requirements, rather than through standard accounts. The reports from ENISA and CERT Polska suggest the models help speed up security analysis, while the researchers say human testing and judgment are still needed.