we_are_coded.by CODE · The world, decoded
БГ

Concept

AI in cyber defense: Daybreak and the defenders

Daybreak is OpenAI's program that gives vetted defenders access to cyber models for authorized work. The model speeds up the hunt for flaws, and the human decides what is real and what to do about it.

Updated on 19 November 2026 · Checked on 19 November 2026

In short: Daybreak is OpenAI's program that gives vetted defenders access to cyber models for authorized work. The model searches and suggests. The human checks and decides.

Daybreak is not a new antivirus. According to OpenAI, it is access to more capable models for people who are trusted: to review software, investigate suspicious activity, confirm whether a vulnerability is real, and develop and test a fix. The work is authorized, meaning it is done by institutions that have the right to touch these systems.

Who uses it

According to OpenAI, NATO, through its Communications and Information Agency (NCIA), has used up to 300 billion input and output tokens per month. A token is a piece of text the model reads or writes, so the number measures volume of work, not number of findings. Outside NATO, access goes to ENISA, the EU's agency, and to national agencies in France, Germany and Poland. Ukraine's Ministry of Digital Transformation uses it for civilian infrastructure. At the 2026 Crossed Swords exercise run by CCDCOE, there was a stream where participants built and evaluated AI agents for cyber operations, with an emphasis on human oversight.

What they found

According to OpenAI, ENISA used the models to identify four vulnerabilities in software used across EU institutions, and fixes are published for all four. A November joint report from ENISA and CERT-EU describes a review of 27 software projects, in which 95% of 88 potential issues assessed by maintainers were confirmed as genuine. CERT Polska used the models to help identify six vulnerabilities in MikroTik RouterOS and coordinated disclosure with the vendor. MikroTik released fixes, and CERT Polska confirmed they stopped the attacks it had observed.

The CERT Polska case is the most telling one. Researchers directed the work, tested findings on real systems and verified their impact. The model sped up the analysis. That is the difference between a tool and a replacement. Even at 95%, a small share of the signals is not real, and someone has to filter it out.

The figures about findings are OpenAI's. The source offers no independent check. One thing is clear from the exercise itself: the agent sits next to the operator, and oversight stays with the human.

The visual is generated code-art, no third-party images.