Source: ENISA
The European Union Agency for Cybersecurity, ENISA, has published a joint advisory with CERT-EU on the use of frontier AI models for security code review. The advisory, published on 12 November 2026 under the EU action plan on cybersecurity and artificial intelligence, reports an experiment in which AI models reviewed the source code of open-source software maintained by EU institutions, bodies, offices and agencies. The scans ran in August 2026 in an isolated container environment, using source code only. No live applications were tested.
What the experiment found
Across 27 codebases, 25 of them already open source and 2 about to become open source, the AI-assisted reviews reported 220 potential security issues. By the time the advisory was published, the code owners had assessed 88 of those findings and confirmed 95 percent of them as true positives, with most of the remainder judged inapplicable because the tools lacked runtime context. The experiment also led to four new CVE identifiers being assigned, which shows the findings were not purely theoretical. ENISA had structured access to frontier models with advanced cybersecurity capabilities, including OpenAI’s GPT-5.5 Cyber, GPT-5.6 Cyber and GPT-6 Astra, and Anthropic’s Mythos 5, run through client tools such as Codex and OpenCode as well as ENISA-developed skills.
AI overrates severity, and triage is the bottleneck
The most practical finding for engineering teams is that the AI tools consistently rated severity higher than the code owners did. The tools classified 42 of the 88 assessed findings as critical or high, while the code owners ended up rating 22 as high and none as critical. Severity usually cannot be determined from source code alone, because it depends on reachability, privileges, trusted components and accepted risks that only the deployment can reveal. The advisory also notes that only around 40 percent of the findings had been assessed by publication, and it names expert triage time, not discovery speed, as the constraint. CERT-EU’s own harness Vigilo, which orchestrates specialised agents focused on vulnerability classes such as injection, authentication and cryptography, and challenges every finding with an adversarial critique step, kept its severity ratings in 74 percent of cases against 60 percent for general-purpose coding agents. CERT-EU has now released Vigilo as open source, so organisations can run it on their own infrastructure with cloud-based or on-premises models.
The cost question
ENISA estimates all the scans together would have cost about EUR 6,649 at commercial rates, or roughly EUR 132 per 10,000 lines of code, consuming more than 611 million tokens. Those figures are indicative, since ENISA had free access to the models, but they point at a genuine shift in the economics of security review: an automated first pass over a codebase now costs less than a fraction of a senior engineer’s day. The catch is what happens next. Someone qualified still has to validate, prioritise and fix what the models find, and that work does not yet accelerate the way discovery does. The advisory recommends running these reviews in secure, isolated environments, with network segmentation, access control and monitoring around agents that can read repositories, create files, execute commands and call remote services.
For European software teams, the actionable conclusion is that AI-assisted code review has moved from hype to documented EU-level evidence, but the value only materialises if you have the human process to absorb the output. If you want help building an AI-assisted security review pipeline with proper isolation and expert triage, contact Excello Digital and let us design a workflow that fits your codebase and your compliance obligations.
