Published
- 4 min read
By Allan D - Editor, AI Security Wire
ENISA and CERT-EU: AI Code Review Flagged 220 Issues, 95% of 88 Checked Held Up
ENISA, the EU’s cybersecurity agency, and CERT-EU, which serves the EU’s own institutions, have published a joint advisory on an experiment in AI-assisted code review. The headline is encouraging: of the findings code owners had assessed, 95 percent were real. The more useful part is everything around that number.
What they did
ENISA has run AI-assisted source reviews since June 2026. The experiment covered 27 codebases (25 already open source, 2 about to be). These were applications deployed in EU entities’ infrastructure or underpinning services those entities rely on. The models had access through vendor programmes: OpenAI’s GPT-5.5 and GPT-5.6 families, a GPT-6 Astra model via Daybreak, and Anthropic’s Mythos 5 via Project Glasswing.
Tooling varied. Some runs used general coding agents (OpenAI’s Codex client with its Codex Security plug-in, and OpenCode) with security skills. Others used Vigilo, a multi-agent harness CERT-EU built and released as open source in September 2026. Vigilo splits work between agents by vulnerability class and runs an adversarial critique step over every finding before it reports. ENISA also wrote two agent skills of its own, Code Auditor and Findings Browser Report.
Everything ran in a sandbox. The models saw source code only, with no production configuration or infrastructure details.
The numbers
The models flagged 220 potential issues. Rated by the AI tools, 4 percent were critical, 44 percent high, 34 percent medium and 18 percent low.
By publication, code owners had finished assessing 88 of them, 40 percent. Of those, 95 percent were true positives. ENISA says the false positives or not-applicable findings were mostly cases where the model lacked runtime context, which the experiment withheld on purpose. Four new CVE IDs came out of it, all assigned to one scanned codebase, which the report does not name.
Severity is where it got messy. After human review, the confirmed findings were rated 26 percent high, 34 percent medium, 29 percent low and 11 percent informational. (The report’s figure shows 4 of the 88 marked false, so those percentages are over the other 84. That 84 is our arithmetic, not a number the report prints.) ENISA’s own reading is that the models showed “a slight tendency” to overrate severity, and most changes on review were downgrades. Code owners kept Vigilo’s severity rating in 74 percent of cases, against 60 percent for the general coding agents with a skill or plug-in. The cost: Vigilo was slower and more expensive.
On money, ENISA estimates the scans cost about EUR 6,648.90 in total, roughly EUR 132.27 per 10,000 lines of code. Those figures are indicative only, because ENISA got model access free under the providers’ programmes.
The bottleneck is people
The report is blunt that finding weaknesses is getting faster and cheaper, and validating, prioritising and fixing are not speeding up to the same extent. Validating, prioritising and fixing “cannot currently be accelerated to the same extent.” Sixty percent of the findings were still waiting on owners when the report went out. ENISA warns that if discovery continues at this pace, the shortage of experts who can verify AI findings becomes a serious constraint, and that researchers should deduplicate and, as far as possible, verify before they report so owners are not buried in duplicates and false positives.
Its suggested process has seven steps: select or build tailored harnesses and skills and run them in isolation, use reputable cyber-capable models to discover and classify issues, triage with qualified humans validating severity, use controlled proof-of-concept exploits to confirm, report through the proper disclosure channel (CERT-EU for EU entities, the project’s own policy otherwise), generate patches for owners to review, and track token spend.
Where this fits
OpenAI published its own account on 17 November in a post on NATO and allied cyber defence. It says ENISA used its models to identify four vulnerabilities in software used across EU institutions, with fixes published, and lists four CVE IDs. We checked those four IDs (CVE-2026-73432, CVE-2026-73431, CVE-2026-73405 and CVE-2026-73374): the CVE records, published 12 August 2026, all name Vulnerability-Lookup as the affected product and carry CVSS v4.0 scores from CIRCL as CNA of 5.1, 8.8, 5.3 and 6.1 (NVD shows no score of its own for them). The records credit “EUVD @ ENISA Team” as reporter. The ENISA report itself does not name the codebase, so tying these IDs to the experiment rests on OpenAI’s post and those credits.
Caveats: this is a self-reported experiment from the agencies that ran it, on a limited set of codebases, with 60 percent of findings unassessed. It also does not say how many findings were fixed. For the offensive side of the same capability curve, see our write-up of the GPT-6.2 system card.
References
Frequently Asked Questions
- What did ENISA and CERT-EU actually test?
- They ran frontier AI models, in several combinations of clients, plug-ins and skills, over the source code of 27 codebases (25 already open source, 2 about to be) maintained by EU institutions, bodies, offices and agencies. The reviews were source-only, with no live applications tested. The models found 220 potential issues, which were passed to code owners.
- Does the 95 percent figure mean 95 percent of all findings are real?
- No. It applies to the 88 findings code owners had assessed by publication, which is 40 percent of the 220. The other 60 percent were still unverified. The report itself calls it a high true-positive rate for the limited subset that got feedback.
- What should teams copy from this?
- The report's own advice is to run AI audit tools in an isolated environment, have qualified humans validate severity, deduplicate before reporting, and use the right coordinated disclosure channel. It also suggests a general coding agent with a security skill as a baseline on pull requests, and a specialised harness for deeper audits.