Home→AI News→NATO and European agencies use AI to find software flaws
AI News

NATO and European agencies use AI to find software flaws

NATO is using OpenAI models to review software, investigate suspicious activity and validate vulnerabilities across allied networks, with use by core institutions reaching up to 300 billion input and output tokens a month, OpenAI said. The alliance is also assessing the costs and practical limits of expanding adoption.

NATO and European agencies use AI to find software flaws

The work involves NATO’s Communications and Information Agency (NCIA), which provides technology and cybersecurity services to the alliance. OpenAI says its models can help security teams develop and test fixes, but has not disclosed how many threats or vulnerabilities NATO has identified with them. The capabilities are offered through Daybreak, OpenAI’s cybersecurity initiative for authorised vulnerability research and remediation.

NCIA General Manager Dr Dylan Browne said AI could speed up vulnerability discovery and help detect emerging threats. He said NATO would assess token consumption, usage patterns and the economics of AI models as adoption grows, helping the alliance manage costs, compare solutions and make procurement decisions.

AI agents were also tested at Crossed Swords 2026, a NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE) exercise held in Tallinn, Estonia, from November 2 to 6. More than 250 participants from over 40 NATO and non-NATO countries took part. In a new training stream, participants designed, deployed and evaluated agents during simulated cyber operations, working alongside human operators. The exercise used locally hosted models and OpenAI Daybreak for the first time, with human oversight retained. Exercise Director Jozef Kostelansky said putting AI on both sides of simulated operations added complexity to coordinating cyber activity with other military operations. Scenarios included offensive and defensive operations and critical infrastructure protection with civilian organisations.

European agencies have reported separate AI-assisted vulnerability investigations. OpenAI said the European Union Agency for Cybersecurity (ENISA) used its models to identify four flaws in software used across EU institutions; fixes were published for all four. In joint research, ENISA and CERT-EU reviewed 27 codebases and reported 220 potential issues to maintainers. By publication in November, maintainers had assessed 88 and confirmed 95% as genuine vulnerabilities. The other 132 remained unassessed, and four CVE identifiers had been assigned. Human reviewers verified findings and severity; some AI severity ratings were higher than maintainers’ manual assessments.

Poland’s CERT Polska said AI models helped identify six vulnerabilities in MikroTik RouterOS. Researchers verified the findings with repeatable experiments before coordinating disclosure with MikroTik. Two flaws could be combined to give attackers administrative access without valid credentials. CERT Polska found evidence suggesting the combination had been exploited before security updates were released in September, including unauthorised account creation and attempts to reach vulnerable devices through SSH. The agency said AI automated parts of its research, but testing and deploying updates and investigating potentially compromised devices still took time.

Comments (0)

Leave a comment

No comments yet. Be the first!