The honest view from inside technical leadership
AI News

AI News

Daily AI news with an engineering leadership take, each story credited to its original AI News report.

Security · AI News

ENISA’s AI Code Review Found 220 Issues in EU Software, and Code Owners Had Checked Only 88 Three Months Later

Europe’s cyber agency measured what frontier models cost to find bugs. The slower, unpriced step was people confirming them, and NATO is now scaling the same class of models.

Line-art illustration of a magnifying glass over lines of code beside a tall stack of flagged findings and a short stack of checked ones, representing the ENISA AI code review triage backlog

An AI code review run by ENISA, the EU’s cybersecurity agency, flagged 220 potential security issues across 27 codebases maintained by EU institutions for an estimated €6,649 in model costs. The findings went to code owners in August 2026, and by the time ENISA and CERT-EU published their joint advisory on November 12, the owners had finished assessing 88 of them. Finding bugs with frontier models is now cheap and fast. Confirming them still runs at the speed of the people who own the code, and that is the constraint engineering leaders need to plan for.

ENISA’s AI Code Review Found 220 Issues for About €6,649

ENISA ran the experiment with CERT-EU, the cybersecurity service for EU institutions, under the EU action plan on cybersecurity and AI. The scans took place in August 2026 and covered 25 open-source codebases plus two due to be open-sourced. The models saw source code only. ENISA withheld configuration and infrastructure details on purpose, and no live applications were tested. The models were OpenAI’s GPT-5.5 Cyber, GPT-5.6 Cyber and GPT-6 Astra through OpenAI’s Daybreak programme, and Anthropic’s Mythos 5 through Project Glasswing. They ran inside OpenAI’s Codex client and its Codex Security plug-in, the OpenCode agent, two ENISA-written agent skills, and Vigilo, a multi-agent penetration-testing pipeline that CERT-EU open-sourced in September 2026. Everything ran in a sandboxed container cut off from other systems. The whole exercise consumed 611,486,042 tokens. ENISA puts the cost at about €6,648.90, or €132.27 per 10,000 lines of code, and calls both figures indicative because the providers gave it model access free of charge.

Code Owners Had Verified 40% of the AI Findings by Publication

Of the 220 findings, code owners had completed their assessment of 88 when the report went out, and the other 132 were still waiting. Of the 88, 95% were confirmed as true positives. ENISA traces most of the rest to the runtime context it had deliberately kept from the models. Owner feedback led to four new CVE identifiers, all assigned to one of the scanned codebases. ENISA’s conclusion is blunt: AI finds weaknesses quickly, while validating, prioritising and fixing them “cannot currently be accelerated to the same extent.” It warns that a shortage of experts able to verify AI findings will become a more serious constraint if discovery keeps this pace.

The AI Models Rated Severity Higher Than the People Who Own the Code

The severity numbers show where human review earns its time. Among the 88 assessed findings, the AI tools rated 3 as critical and 39 as high. After review, the code owners rated 22 as high, none as critical, 9 as informational, and 4 as false. Most changes were downgrades. The type of tool mattered. Owners kept the severity Vigilo assigned in 74% of cases, against 60% for general-purpose coding agents with a security skill or plug-in. ENISA credits Vigilo’s adversarial critique step, which re-examines each finding and merges duplicates before the report is written. That step costs more time and more tokens, which is why ENISA suggests the specialised harness for in-depth audits and the cheaper general-purpose agent as a baseline check on pull requests.

NATO Is Scaling the Same Models and Says It Will Track the Economics

On November 17, OpenAI said NATO’s Communications and Information Agency (NCIA) gives NATO’s core institutions access to its cybersecurity models, with use reaching up to 300 billion raw input and output tokens a month. OpenAI also said its Daybreak models were part of the technical environment for Crossed Swords 2026, the exercise run by the NATO Cooperative Cyber Defence Centre of Excellence, in a new training stream where participants built and evaluated AI agents for cyber operations. OpenAI added that ENISA and national agencies in France, Germany and Poland have access to its cyber models. Dylan Browne, NCIA’s general manager, said in OpenAI’s announcement that NATO will build an understanding of token consumption and model economics to optimise costs and inform procurement. Neither OpenAI nor NATO has said how many vulnerabilities NATO has found or what the access costs. Muhammad Zulhusni’s AI News report brought the NATO and ENISA threads together on November 20.

Why This Matters for Engineering Leadership

ENISA published the token bill for its scans. The figure it could not publish was the engineering time needed to clear 220 findings, and three months in, that time had covered 40% of them. Any team pointing frontier models at its own repositories will hit the same wall, because the model’s output lands on the same few engineers who already own those services. Plan the triage side before you scale the scanning side. Give every repository in scope a named owner and a triage deadline before the first scan runs. Deduplicate and reproduce findings centrally before they reach that owner, as ENISA recommends, so the owner sees a short list of confirmed issues. Treat AI severity as provisional: in ENISA’s data, 42 findings the models called critical or high became 22 highs after review, so paging someone on an AI-rated critical will burn trust fast. Give the model deployment context where you can, including reachability, privileges and existing controls, since missing context drove most of ENISA’s false positives. Then measure the pipeline the way NATO says it will measure tokens. Track findings assessed per week, the age of open AI findings, the confirmation rate and the downgrade rate. If the backlog grows faster than the assessed count, slow the scans down until your people can keep up.

FAQ

What did the ENISA and CERT-EU AI code review find?

Frontier AI models flagged 220 potential security issues across 27 codebases maintained by EU institutions. By publication on November 12, 2026, code owners had assessed 88 and confirmed 95% of those as true positives.

How much did the ENISA AI code review cost?

ENISA estimates about €6,649 for all scans, or roughly €132 per 10,000 lines of code, using 611 million tokens. It calls the figures indicative because model providers gave it access free of charge.

Which AI models did ENISA use?

ENISA used OpenAI GPT-5.5 Cyber, GPT-5.6 Cyber and GPT-6 Astra through OpenAI’s Daybreak programme, and Anthropic Mythos 5 through Project Glasswing.

Did the AI models rate severity accurately?

They tended to rate it higher. Of the 88 assessed findings, the models called 42 critical or high, while code owners rated 22 as high and none as critical.

How is NATO using AI for cybersecurity?

OpenAI says NATO’s Communications and Information Agency gives NATO’s core institutions access to its cyber models, with use reaching up to 300 billion tokens a month. NATO has not disclosed what the models found.

What should engineering teams do before running AI security scans at scale?

Assign an owner and a triage deadline for every repository first, deduplicate and pre-validate findings before they reach that owner, and treat AI severity ratings as provisional until a person with deployment context confirms them.

Originally reported by Muhammad Zulhusni for AI News, November 20, 2026: How NATO uses AI for cybersecurity and threat detection. Primary source: ENISA and CERT-EU, “Joint Advisory on the Use of Frontier AI Models for Code Review” (November 12, 2026) (PDF); OpenAI, “Strengthening NATO and allied cyber defenses with AI” (November 17, 2026).

Read more AI News