NATO uses OpenAI models to find vulnerabilities monthly
The alliance is scaling its use of large language models to scan software and defend civilian infrastructure across the EU and Allied nations.

The short version
- NATO processes up to 300 billion raw input and output tokens every month using OpenAI models.
- The EU's cyber agency, ENISA, used these models to find four vulnerabilities in software used by EU institutions.
- Ukraine's Ministry of Digital Transformation uses the tools to protect civilian infrastructure.
- Researchers from CERT Polska used the models to identify six vulnerabilities in MikroTik RouterOS.
- NATO's 2026 Crossed Swords exercise focused on building AI agents for cyber operations with human oversight.
Why it matters
This shows that large-scale government entities are moving from experimenting with AI to integrating it into core defense workflows. It demonstrates how automated scanning can identify security risks faster than manual review while maintaining human oversight.
Scaling AI for cyber defense
NATO’s Communications and Information Agency (NCIA) has granted its core institutions access to OpenAI’s cybersecurity models. This initiative is designed to protect the networks of Allied nations by identifying security flaws before they can be exploited. The scale of this operation is significant, with the agency processing up to 300 billion raw input and output tokens every month.
The data shows that these models are being used to scan software that European institutions rely on for daily operations. By automating parts of the security review process, the alliance aims to stay ahead of sophisticated cyber threats. This work is part of a broader effort to modernize how digital infrastructure is defended against constant attacks.
The partnership is facilitated through Daybreak, which is OpenAI’s specific program for vetted defenders. This program allows authorized users to perform tasks like reviewing code, chasing suspicious activity, and testing various security fixes. It ensures that the high-powered models are used in a controlled environment for official government work.
Real-world vulnerability discovery
The practical application of these models has already resulted in the discovery of multiple security flaws. As helpnetsecurity.com reported, the EU’s cyber agency, ENISA, utilized these models to locate four flaws in software used by EU institutions. All four of these issues have since been addressed with released fixes.
Other agencies have seen similar success in identifying specific technical weaknesses. For example, CERT Polska used OpenAI models to assist in finding six vulnerabilities in MikroTik RouterOS, which is the operating system used on MikroTik network gear. Researchers from the group directed the AI's work and then verified the findings on real systems to ensure accuracy.
The results from these tests have been highly effective at identifying genuine problems. During a review of 27 software projects, maintainers assessed 88 potential issues and confirmed that 95% of them were real. This high accuracy rate suggests that AI can be a reliable tool for triaging security concerns at a massive scale.
Geographic reach and infrastructure
The reach of this AI-driven defense extends across several different countries and types of infrastructure. Ukraine’s Ministry of Digital Transformation currently uses the Daybreak program to defend its civilian infrastructure against cyber threats. This highlights the use of the technology in active conflict zones and for protecting public services.
Several other national agencies have been granted access to these specific OpenAI models. This includes France’s ANSSI, as well as national agencies in both Germany and Poland. By providing these tools to various countries, NATO is creating a more unified approach to cybersecurity across the alliance.
The focus remains on protecting the software that underpins modern governance and public utility. Because these systems are used every day by millions of people, the ability to scan them for vulnerabilities is a priority. The use of frontier models allows for a more comprehensive review of complex codebases than traditional methods might allow.
Human-led AI operations
While the technology is powerful, NATO is emphasizing the importance of human oversight in cyber operations. During the 2026 Crossed Swords exercise, run by the CCDCOE, participants built and tested AI agents for cyber operations. A primary focus of this exercise was ensuring that humans remain in charge of the AI's actions.
The goal is to create a system where AI agents can perform the heavy lifting of scanning and analysis while humans make the final decisions. This approach seeks to balance the speed of machine processing with the nuanced judgment of human operators. It prevents the risks associated with fully autonomous cyber defense systems.
As the alliance continues to scale its adoption of these tools, it is looking to understand the underlying economics of the models. Dr. Dylan Browne, the NCIA general manager, stated that they will build a deeper understanding of token consumption and usage patterns. This data will help the alliance optimize costs and make better procurement choices in the future.
What to watch
- How NATO will evaluate competing AI models against OpenAI's performance.
- Whether more civilian infrastructure in other countries will adopt similar AI scanning tools.
Try it
Look for reports on 'Daybreak' by OpenAI to see how other vetted organizations are using AI for authorized security work.
Sources and links
- Officialopenai.com/index/strengthening-nato-allied-cyber-defences-ai/ openai.com
- Reportinghelpnetsecurity.com: the article we sourced
First reported by helpnetsecurity.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.