News

OpenAI says NATO institutions are using its cyber models

OpenAI says NATO institutions and allied agencies are using its cyber models. Official records independently confirm a bounded NATO exercise and a supervised Polish vulnerability project.

D
Nov 17, 2026 · 2 min read

OpenAI said on November 17 that it is working with the NATO Communications and Information Agency to give NATO core institutions access to advanced cybersecurity models for protecting Allied networks. It also said NATO has processed as many as 300 billion raw input and output tokens in a month through the work. No opened NATO or NCIA publication independently confirms the deployment’s scope or the usage figure.

A separate account from the NATO Cooperative Cyber Defence Centre of Excellence confirms a narrower use of OpenAI Daybreak during Crossed Swords 2026. The five-day exercise ran in Tallinn from November 2 through November 6, bringing together more than 250 participants from over 40 NATO and non-NATO countries.

According to CCDCOE, the exercise introduced an AI-agent training stream in which participants designed, deployed and evaluated agents. The technical environment combined on-premises AI with Daybreak, while operators maintained human oversight across offensive and defensive scenarios. The record confirms Daybreak’s presence in a bounded training environment, not routine use on live NATO networks or access across NATO core institutions.

OpenAI’s announcement also says cyber agencies in France, Germany and Poland, along with the European Union Agency for Cybersecurity, have received access to its cyber models. The company says Ukraine’s digital-transformation ministry is using Daybreak to defend civilian infrastructure. Those access and deployment claims were not independently confirmed in the opened evidence.

There is separate support for one allied example. CERT Polska says its researchers used GPT-5.5-cyber and GPT-5.6-sol through an OpenAI government and trust program while identifying and coordinating disclosure of six MikroTik RouterOS vulnerabilities. The controlled-access arrangement provides context for OpenAI’s release of GPT-6 Cyber to vetted security researchers. CERT Polska said the models automated laboratory work and hypothesis exploration; researchers chose and supervised the work.

Before confirming the findings, CERT Polska required reproduction on real systems, negative controls, clean-state tests and human assessment. It said MikroTik released fixes for all six flaws and that fully updated versions prevented the attacks it had observed. Two software branches needed another update after their initial fixes proved incomplete.

The available records do not disclose procurement terms, deployment architecture, data-handling controls, audit procedures or human-approval gates outside the exercise. NATO’s public cyber-defense overview says the Alliance focuses on protecting its own networks and helping Allies strengthen their national resilience.

More news