HomeInsightsENISA Finds 220 Potential Software Security Issues Using Frontier AI Models
TechnologyNov 13, 2026, 11:59 AM

ENISA Finds 220 Potential Software Security Issues Using Frontier AI Models

Kosmos Consult

Published by

Kosmos Consult3 min read

A joint ENISA and CERT-EU experiment used frontier AI models to review 27 open-source codebases maintained by EU institutions. The systems identified 220 potential issues, and 95% of the 88 findings assessed so far were confirmed as genuine security problems.

Cybersecurity vulnerability database graphic illustrating the software security review conducted by ENISA and CERT-EU.

Cybersecurity vulnerability database graphic illustrating the software security review conducted by ENISA and CERT-EU.

On this page

AI is becoming another tool for finding weaknesses in software

Software security reviews often require experienced engineers to examine large codebases, trace how components interact and identify weaknesses that could expose systems to attackers.

Artificial intelligence may help shorten that process.

A new experiment by the European Union Agency for Cybersecurity (ENISA), conducted with support from CERT-EU, examined whether frontier AI models could assist with security reviews of open-source software maintained by EU institutions.

The findings were published on November 13, 2026. Across 27 codebases, the models identified 220 potential security issues.

What the experiment examined

The review covered 25 existing open-source codebases and two that were expected to become open source.

Researchers examined source code rather than live applications. The models were not given information about production deployments, and the experiment did not involve testing running services.

ENISA used Vigilo, a specialised security testing framework developed by CERT-EU, to support the work.

The identified issues were shared with the relevant software owners for assessment.

How many findings were confirmed?

At the time of publication, software owners had completed assessments of 88 of the 220 reported findings.

ENISA said 95% of those reviewed findings were confirmed as genuine security issues. Four new Common Vulnerabilities and Exposures identifiers had also been assigned.

The distinction matters. The 95% figure applies to the subset that had been assessed, not all 220 findings.

The remaining reports were still undergoing review, and identifying a potential issue does not automatically mean that it can be exploited in a real-world deployment.

What this means for software developers

The experiment offers evidence that AI models can assist security teams with code auditing and vulnerability discovery.

Developers may use these tools to review large repositories, identify suspicious patterns and prioritise areas that deserve closer human examination.

But AI-generated findings still require validation. A model may misunderstand how a function is used, overlook an important dependency or report a weakness that cannot be reproduced.

Security teams must also examine the context in which a vulnerability appears, determine its practical impact and verify that a proposed fix does not introduce another problem.

ENISA's experiment is not proof that AI can replace professional security researchers. It shows that these systems can contribute useful findings when placed inside a structured review process with expert oversight.

For organisations maintaining critical software, the opportunity is to combine automated analysis with human expertise, making security reviews more thorough without treating model output as a final verdict.

Sources: ENISA report reproduced by INSIGHT EU Monitoring, November 13 · CERT-EU Vigilo repository · European Union Agency for Cybersecurity

Similar insights