European cyber defense relies more and more on American artificial intelligence. THE November 17, 2026, OpenAI announced the opening of Daybreakits access program to its specialized cybersecurity models,NATO and to several European governments. In France, theANSSIthe National Agency for Information Systems Security, is among the beneficiaries, as are the national agencies of Germany and Poland. An announcement which illustrates the growing effectiveness of these tools, but also Europe’s dependence on a foreign supplier for such a sensitive mission.
What Daybreak is
Daybreak is the program through which OpenAI provides access to its most powerful cybersecurity models to verified defendersfor authorized work: reviewing code for flaws, investigating suspicious activity, testing fixes. These capabilities are considered too sensitive to be open to everyone, because the same tools that spot a flaw to fix it could help an attacker exploit it.
OpenAI is increasingly strictly regulating its most advanced models in this area, particularly since it deemed its Astra model “critical” in cybersecurity and limited its initial access.
Who has access to it
| Organization | Declared use |
|---|---|
| NATO Communications and Information Agency (NCIA) | Protection of the networks of the central institutions of the Alliance |
| ANSSI (France) | Access granted, uses not detailed |
| National agencies of Germany and Poland | Access granted |
| Ukrainian Ministry of Digital Transformation | Defense of civil infrastructure |
| ENISA and CERT-EU | Analysis of software security of European institutions |
L’NATO is the most intensive user. According to Help Net Security, its communications and information agency used up to 300 billion tokens per montha unit that measures the volume of text processed by the model. Its director, Dylan Browne, speaks of a “generational leap” for cyber defense. The agency says it wants, as usage develops, to better understand the consumption and economics of these models to optimize its costs and evaluate competing solutions.
When exercising Crossed Swords 2026organized by the NATO Center of Excellence for Cooperative Cyber Defense, participants deployed for the first time AI agents for cyber operations, under human supervision.
Concrete results in Europe
The most telling figures come fromENISAthe European Union agency for cybersecurity, and the CERT-EUthe team that protects the systems of the European institutions. Together, they analyzed 27 software projects using several cutting-edge AI models, including those from OpenAI. Those responsible for this software have examined 88 potential problems reported by AI and confirmed 95% like real faults.
Among the results, four vulnerabilities in software used by European institutions, all since corrected. These flaws had already been revealed by Politico in September 2026. One of them, deemed high risk with a rating of 8.8 out of 10, affected Vulnerability-Lookup, free software for tracking vulnerabilities. It would have allowed an attacker to take control of accounts: the software did not check whether an activation or recovery link had already been used, which allowed it to be reused to reset a password, again and again, as long as the link remained valid.
This example clearly illustrates what AI knows how to do: identify logical errors that a human proofreader can miss, because they do not arise from a faulty line of code but from an absent verification. The model examines the code, formulates hypotheses for flaws, then experts check and classify them.
In Poland, the CERT Polandthe national incident response team, found with the help of OpenAI models six flaws in RouterOS, the network equipment system from the manufacturer MikroTik. Two of them, combined in an attack that the team called MikroTrick, had been actively exploited since at least early September 2026: they gave total control of routers whose remote access by SSH was open on the internet. MikroTik has released patches, which CERT Polska has confirmed blocks the observed attacks.
Why it matters
Time is against the defenders. According to ENISA, an attacker can now weaponize a vulnerability in 15 minutes after its public disclosure. Faced with this speed, the ability to analyze thousands of lines of code in a few minutes becomes a decisive advantage.
The issue goes beyond institutions. Vulnerabilities discovered in widely used free software benefit all users, including businesses and individuals, when fixed. Conversely, the same tools in the wrong hands could accelerate the discovery of vulnerabilities by attackers: this is the ambivalence of these models, and the reason why their access is controlled.
For French companies, the lesson is practical. The time between the publication of a patch and its exploitation is reduced: quickly apply security updates becomes a top priority, especially for equipment exposed to the internet, such as routers. According to Alliancy, the deputies have also just voted on the bill which transposes the European directive NIS 2, which must require more than 18,000 entities to strengthen their cybersecurity by the end of 2028.
The question of sovereignty
That national security agencies rely on the models of an American company raises an obvious question of dependence. The analyzed data, the institutional code and the details of the flaws pass through tools whose design or access conditions Europe has no control over, which are subject to change.
ENISA is aware of this. In a paper on cutting-edge AI, the agency said European organizations not only need access to these models, but also develop their own. The Commission published an Action Plan on AI and Cybersecurity in July 2026, under which ENISA and the Joint Research Center are building a secure testing platform for the most advanced models. European players are starting to position themselves: OVHcloud is planning a version of its future Olala models specializing in cybersecurity.
A race between defenders and attackers
The opening of Daybreak marks a milestone: AI is no longer an experimental tool for cybersecurity teams, but an instrument used on a large scale by military alliances and national agencies. The results obtained by ENISA and CERT Polska show its effectiveness, provided that experts are kept to verify and decide. It remains to be seen whether Europe will manage to develop its own tools before its dependence becomes another vulnerability. Our cybersecurity section will follow these developments.

