OpenAI opens Daybreak to NATO and the EU: 95% of 88 reported bugs turned out to be real
OpenAI has opened Daybreak to NATO and European governments. The AI reads code, looks for vulnerabilities and helps close them. In an ENISA and CERT-EU report covering 27 projects, 95% of 88 reported issues were real bugs, not false alarms. That means specialists won't have to drown in junk warnings. For security teams, AI code auditing is becoming a working tool, but people still make the decisions.
AI-processed from The Decoder; edited by Hamidun News
OpenAI has opened access to its cyber defense tool Daybreak for NATO and several European governments. Daybreak is a service in which OpenAI models read program code, look for vulnerabilities (holes that can be used to break into a system) and help write patches. The headline number is 95% of 88: that is the share of reports, submitted after reviewing 27 software projects, that turned out to be real bugs rather than false positives. This matters in practice because conventional security scanners produce hundreds of "findings," half of them junk, and someone has to sort through them by hand. Here, almost every finding was worth fixing.
What has already been found
A telling example is the joint report from ENISA, the European cybersecurity agency, and CERT-EU, the computer incident response team for EU institutions. They examined 27 software projects and reported 88 issues. According to the report, 95% of them turned out to be real bugs, or about 84 cases. Separately, ENISA used these models to find four vulnerabilities in software used by EU institutions.
The Polish team CERT Polska found six vulnerabilities in MikroTik routers, and attackers were already actively exploiting them. These are not theoretical holes but problems that were used in real attacks. Ukraine, for its part, uses Daybreak to protect civilian infrastructure. France, Germany and Poland have also received access to the tool.
Humans stay in the loop
The NATO Communications and Information Agency (NCIA) says it is already actively using foundation models. Its head, Dylan Brown, called the shift a "generational leap" in cyber defense. That is the assessment of an interested user, so it should be read as an opinion, not a measurement.
A more tangible fact comes from the NATO exercise Crossed Swords 2026, where participants deployed AI agents for cyber operations under human control for the first time. The agent works as a fast assistant: it combs through code and suggests options, while a specialist decides on the patch or action. For government organizations, where mistakes are costly, this setup looks the most realistic.
What this means for developers
First, AI code auditing is no longer a hobbyist experiment. If EU institutions and a defense alliance use it, the question shifts from "can we trust it" to "how do we build it into our process." A team that ships software should decide in advance who reviews such reports and how quickly fixes go out.
Second, a growing number of discovered vulnerabilities increases the load on the people who fix them. Maintainers of popular libraries and small vendors, such as router makers, may get a flood of reports faster than they can respond. It helps to prepare a clear channel for security reports and a response template.
Third, anyone selling software to the public sector should recognize that customers in Europe already work with AI defense tools. The question of how your product will hold up under such an audit will come up more and more often.
Limitations
The data comes from materials by OpenAI and its partners, and the news includes no independent verification of the methodology. The 95% figure speaks to the precision of what was found, not to completeness: it is unknown how many real bugs the models missed. The terms of access, pricing and technical details of how Daybreak works have not been disclosed. Finally, the same capabilities could in theory help attackers too, so access control remains an open question. One thing is clear for now: AI vulnerability hunting has moved from demonstrations into the practice of government teams.
- Nov 20, 2026OpenAI Agent Wrote a Script to Secretly Copy Email Instead of a Health Document
- Nov 20, 2026Proton releases Lumo Desktop: an AI agent with file and shell access, under your control
- Nov 20, 2026Linear Hits $100M ARR Without SDRs: Agents Now Create Over 50% of Issues, Up From 3%
- Nov 19, 2026Amazon and Google will hold 28.4% of Anthropic after its IPO, but without voting rights
Key Questions Answered
What has already been found?
A telling example is the joint report from ENISA, the European cybersecurity agency, and CERT-EU, the computer incident response team for EU institutions. They examined 27 software projects and reported 88 issues. According to the report, 95% of them turned out to be real bugs, or about 84 cases. Separately, ENISA used these models to find four vulnerabilities in software used by EU institutions.
What are the key details on Humans stay in the loop?
The NATO Communications and Information Agency (NCIA) says it is already actively using foundation models. Its head, Dylan Brown, called the shift a "generational leap" in cyber defense. That is the assessment of an interested user, so it should be read as an opinion, not a measurement. A more tangible fact comes from the NATO exercise Crossed Swords 2026, where participants deployed AI agents for cyber operations under human control for the first time.
What this means for developers?
First, AI code auditing is no longer a hobbyist experiment. If EU institutions and a defense alliance use it, the question shifts from "can we trust it" to "how do we build it into our process." A team that ships software should decide in advance who reviews such reports and how quickly fixes go out. Second, a growing number of discovered vulnerabilities increases the load on the people who fix them.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.