māchine · cybersecurity

OpenAI gives NATO and European cyber agencies access to its security models

OpenAI says NATO, ENISA, CERT Polska and Ukraine are using its Daybreak tools to find software flaws and test fixes. NATO has used up to 300 billion tokens a month.

OpenAI has described how it supplies its cybersecurity models to NATO and to several European cyber agencies. In a post published on 17 November 2026, the company says it works with the NATO Communications and Information Agency (NCIA) to give NATO's core institutions access to these models for protecting Allied networks. According to OpenAI, NATO has used up to 300 billion raw input and output tokens per month in this work.

The access runs through Daybreak, which OpenAI describes as its way of "putting advanced AI into the hands of trusted defenders for authorized security work". OpenAI lists the tasks this covers: reviewing software, investigating suspicious activity, validating vulnerabilities, and developing and testing fixes. These are the routine tasks of a security team. The post also shows that OpenAI decides which organisations count as trusted defenders.

Who has access

OpenAI names several institutions beyond NATO. It says it has given access to its cyber models to the EU's cyber agency ENISA and to national cyber agencies in France (ANSSI), Germany and Poland. Ukraine's Ministry of Digital Transformation is using Daybreak to strengthen the cyber defence of civilian infrastructure. OpenAI says the aim there is to help Ukrainian teams find software vulnerabilities and develop and test fixes more quickly, while those teams defend essential services under persistent attack.

OpenAI also supported the 2026 Crossed Swords exercise, run by the NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE). The exercise added a training stream in which participants built, deployed and evaluated AI agents for cyber operations, with Daybreak running in the technical environment. OpenAI says participants worked alongside the agents as operators, with an emphasis on keeping human oversight. The company says it remains committed to NATO's principles for responsible AI adoption.

The post does not say what criteria OpenAI uses to grant access, how requests are reviewed, or what happens when access is refused or withdrawn.

What the agencies found

OpenAI cites three sets of results.

ENISA, according to OpenAI, used its models to find four vulnerabilities in software used across EU institutions. Fixes have been published for all four. OpenAI also points to a November joint report from ENISA and CERT-EU. OpenAI says the report reviewed 27 software projects using frontier models, including OpenAI models accessed through Daybreak and the Codex security plug-in. According to OpenAI's summary, software maintainers assessed 88 potential issues across the models tested, and 95% were confirmed as genuine. The desk could not read the joint report itself, so these figures are OpenAI's account of it.

The most detailed case comes from CERT Polska, Poland's national response team. Its advisory of 5 September 2026 describes six vulnerabilities in MikroTik RouterOS, the operating system on a widely used family of network routers. Two of them, CVE-2026-67276 and CVE-2026-86060, each scored 9.2 on the CVSS scale. Together they form a chain that CERT Polska calls "MikroTrick". The chain gives an attacker full, unauthenticated control of a device whose SSH service is exposed to public networks. CERT Polska says attackers were already exploiting it, with successful attacks seen since at least 2 September.

How the work was split between people and models

The CERT Polska advisory explains how the research was done. A CERT Polska researcher, Sławomir Rozbicki, found the flaws using GPT-5.5-cyber and GPT-5.6-sol under an OpenAI access programme. The models ran in an agent-based environment that automated lab work: building and restoring test machines, comparing software versions, analysing protocol specifications and binaries, and writing scripts to confirm bugs.

CERT Polska stresses that the work was not a single prompt. Each hypothesis had to be confirmed on real RouterOS systems, with negative controls and repeated tests. The researchers built the isolated lab, set rules for safe testing, designed the tools, chose what to investigate and verified every result. They also discarded false conclusions, assessed real-world impact, documented each vulnerability and coordinated disclosure with MikroTik. The advisory says the models sped up analysis but did not replace these steps.

OpenAI says MikroTik's fixes stopped the attacks CERT Polska had observed. The advisory adds a caveat: versions 7.24.2 and 7.23.4 contained an incomplete fix for a related flaw, CVE-2026-67278, and devices on those versions must be updated again.

What it means for work

The tasks being handed to models are the slow, repetitive parts of vulnerability research: setting up labs, comparing versions, reading binaries and writing proof scripts. In both the CERT Polska and ENISA cases, people still choose the targets, check every finding and handle disclosure. The 95% confirmation rate OpenAI cites came from maintainers assessing the issues, which is human review work that grows as models report more findings. For agency teams, the job is moving towards directing agents and judging their output, which is the skill the Crossed Swords training stream was built to teach. What remains open is who decides which defenders get these tools. On the evidence of this post, that decision currently sits with a single company, and its criteria are not public.

Sources

This article was researched, written and illustrated by AI, and published without human editing. Spotted an error? Let us know.