NATO is expanding its use of artificial intelligence to identify software vulnerabilities, investigate suspicious activity, and support cyber defence operations across its allied networks. The initiative involves OpenAI and NATO’s Communications and Information Agency (NCIA), which provides technology, communications, and cybersecurity services to the alliance.
According to OpenAI, NATO’s core institutions now have access to advanced AI models for cybersecurity operations, with usage reaching as much as 300 billion tokens per month. The technology is intended to help security teams review software, investigate potential threats, validate vulnerabilities, and develop fixes for identified security issues.
The models are being provided through OpenAI’s broader cybersecurity initiative, Daybreak, which focuses on authorised vulnerability research and remediation. AI is being positioned as a tool to accelerate investigations and help security professionals identify weaknesses before they can be exploited.
However, NATO has not disclosed how many vulnerabilities or security incidents have been identified through the deployment, leaving the overall effectiveness of the initiative difficult to quantify.
NATO evaluates AI costs and cybersecurity capabilities
NCIA is also evaluating the operational and financial implications of expanding AI across its cybersecurity operations. This includes understanding model usage, token consumption, operating costs, and how different AI solutions compare as adoption increases.
Dr Dylan Browne, NCIA General Manager, said AI could accelerate vulnerability discovery and improve the detection of emerging threats. He also highlighted the importance of understanding the economics of AI deployment as the organisation expands its use of the technology.
NATO will use these findings to evaluate competing solutions, assess procurement decisions, and determine how AI capabilities can be scaled across its operations.
AI agents tested during NATO cyber exercises
AI agents were also evaluated during Crossed Swords 2026, a cybersecurity exercise organised by the NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE).
The exercise brought together more than 250 participants from over 40 NATO and non-NATO countries. Military cyber operators, digital forensics analysts, threat hunters, and other cybersecurity specialists participated in simulated operational scenarios.
For the first time, the exercise included a dedicated training stream in which participants designed, deployed, and evaluated AI agents within simulated cyber operations. The agents were integrated into existing workflows alongside human operators, allowing participants to examine how AI could support practical cybersecurity tasks.
The technical environment included locally hosted AI models and OpenAI Daybreak, which was introduced into the exercise for the first time. Participants evaluated AI-assisted operations across both offensive and defensive scenarios, including activities involving critical infrastructure protection.
Jozef Kostelansky, Exercise Director of Crossed Swords 2026, said AI was now being incorporated on both sides of simulated cyber operations. He noted that integrating AI agents into operational workflows added complexity to coordinating cyber activities with broader military operations.
AI helps identify vulnerabilities across European software
AI-assisted vulnerability research is also being used by European cybersecurity organisations outside NATO.
The European Union Agency for Cybersecurity (ENISA) reportedly used AI models to identify four vulnerabilities in software used across EU institutions. All four vulnerabilities received fixes, with findings also documented through research involving ENISA and CERT-EU.
In a joint assessment, the agencies examined 27 software codebases maintained by EU institutions and identified 220 potential security issues. These findings were reported to the relevant software maintainers for investigation and validation.
Of the 88 reported issues assessed by maintainers at the time, 95% were confirmed as genuine vulnerabilities. The remaining findings were still awaiting review, while four new Common Vulnerabilities and Exposures (CVE) identifiers had been assigned.
The assessment combined several AI models and security tools, with human reviewers responsible for validating findings and evaluating their severity. Researchers also identified cases where AI-generated severity ratings were higher than the assessments produced by software maintainers following manual investigation.
AI-assisted research uncovers MikroTik vulnerabilities
Poland’s national cybersecurity incident response team, CERT Polska, has also used AI models to support vulnerability research involving MikroTik RouterOS, the operating system used in MikroTik networking equipment.
The research identified six vulnerabilities through AI-assisted analysis of software binaries, comparisons between different software versions, and automated testing. Researchers subsequently verified the findings through repeatable experiments before coordinating their disclosure with MikroTik.
Two of the vulnerabilities could be combined to allow unauthorised administrative access to affected devices without valid credentials. CERT Polska also found evidence suggesting that attackers had exploited the vulnerability combination before MikroTik released security updates in September.
Researchers analysed publicly available attack logs and reports from affected administrators to reconstruct the incidents. The investigation uncovered evidence of unauthorised account creation and attempts to access vulnerable devices through SSH connections.
CERT Polska said AI helped automate parts of the vulnerability research process, particularly software comparisons and testing. However, validating findings, deploying security updates, and investigating potentially compromised systems continued to require significant time and human involvement.
AI becomes part of modern cyber defence
NATO’s adoption of AI, alongside research conducted by European cybersecurity agencies, highlights how AI models are increasingly being incorporated into vulnerability discovery, threat investigation, and incident response.
While the technology can help automate repetitive analysis and accelerate security research, the findings still require human validation before organisations can confidently act on them. False positives, severity misclassification, operational costs, and the complexity of deploying fixes remain important considerations.
As governments and military organisations continue evaluating AI agents in cybersecurity environments, their ability to integrate these tools into existing workflows while maintaining human oversight will likely determine how widely the technology is adopted.
Source: https://www.artificialintelligence-news.com/news/nato-ai-cybersecurity-threat-detection/


