P.K. SHARMA

Cyber security intelligence, AI governance, practitioner analysis

ENISA and CERT-EU: AI review of 27 codebases made 220 findings; 88 assessed, 95% of those confirmed

ENISA and CERT-EU's joint advisory, dated November 2026, says AI models found 220 issues in 27 codebases and code owners had assessed 88 of them at publication. The 95% confirmed rate covers those 88 only, and how many of the other 132 are true is not stated.

By · LinkedIn · · 25 min read

The 95% describes 88 of 220 findings, and the report says so

ENISA and CERT-EU's joint advisory, a 15 page PDF dated November 2026, reports an experiment in which AI models, with harnesses and skills, read the source code of 27 open source codebases and produced 220 findings. By the time of publication code owners had assessed 88 of them, which the report's Figure 3 gives as 40% (derived: 88 / 220 = 40.0%). The headline rate, 95% confirmed as true positives, is a rate over those 88. The other 132 (derived: 220 - 88) had no answer in the report: it says "The remaining findings were still being assessed." The executive summary limits its own claim in the same way: it says ENISA found a high true-positive rate "for the limited subset of AI-identified issues on which code owners gave feedback".

The report's Figure 6 marks 4 of the 88 as false, so the confirmed count is 84 (derived: 88 - 4 = 84), which is 95.5% of 88 (84 / 88; the report rounds to 95%) and 38.2% of 220 (derived). Across all 220 the true count can only be said to lie between 84 and 216 (derived: 84 if none of the 132 unassessed findings is true, 84 + 132 = 216 if all are).

What that does not establish. It does not establish how many of the unassessed 132 are true, or why these 88 were assessed first: the report does not say how they were chosen. It does not establish that AI code review is cheap or expensive, that 95% proves reliability, or that the four CVE ids are typical. This site reports what the report and the records print, works out the ratios, and labels every figure it worked out "derived".

Whose account this is. The report is ENISA and CERT-EU's account of their own experiment. The experiment also tested a harness CERT-EU developed (Vigilo) and two skills ENISA developed, and the report says so: the authors tested their own tools. The report says ENISA was given access to the models free of charge under the providers' access programmes, OpenAI's Daybreak and Anthropic's Project Glasswing. OpenAI makes the GPT models, the Codex client and the Codex Security plug-in in the report's tables; Anthropic makes Mythos 5. Each has an interest in an experiment like this being read as a success, and OpenAI's page of 17 November cites the report in support of its own cyber programme.

What the experiment was

The report says ENISA has been running AI assisted source code reviews since June 2026, under key action 4 of the EU action plan on cybersecurity and AI. ENISA compiled a non exhaustive list of open source software maintained by Union entities, "focusing on applications deployed in Union entities' infrastructure or underpinning essential services"; the selection was agreed with the code owners, and a selected group of Union entities was invited to submit the code of critical, internet facing open source applications. Of the 27 codebases, 25 were already open source and 2 were about to become so.

The set-up is the report's Figure 1: a container, on a Docker engine, on a host machine, holding plug-ins, skills, reporting templates, scripts for analytics and an agentic client. The report says the environment was isolated "so that the AI models could not access any systems, data or services beyond the scope of the experiment". The reviews used source code only and the models were given no information about production deployments, "(e.g. configuration or infrastructure)"; the report says this was deliberate, and that no live applications were tested.

The models and harnesses as the report's Tables 1 and 2 list them, with the access programme and maker as printed. The report says they were used in various combinations and does not say which combination produced which finding.

ItemKind and accessMade by, as printed
OpenAI GPT-5.5 family (including cyber)Model, Daybreak programmeOpenAI
OpenAI GPT-5.6 family (including cyber)Model, Daybreak programmeOpenAI
OpenAI Astra modelModel, Daybreak programmeOpenAI
Anthropic Mythos 5Model, Project GlasswingAnthropic
Codex clientClientOpenAI
Codex Security plug-inPlug-inOpenAI
OpenCode agentic clientClientAnomaly
VigiloClient, white box mode on source code onlyCERT-EU
Code AuditorSkillENISA
ENISA Findings Browser (Findings Browser Report)SkillENISA

The text of the report names the models as "OpenAI GPT-5.5 Cyber, GPT-5.6 Cyber, GPT-6 Astra and Anthropic Mythos 5"; its Table 1 prints the Daybreak programme once, against the three OpenAI rows. Vigilo is described as an AI powered penetration testing pipeline that CERT-EU released as open source in September 2026. It runs several specialised agents and an adversarial critique step that re-examines every finding, challenges its severity and merges duplicates. The report does not say which Vigilo version ran in August, and this site did not read the repository.

Drawn from the report's set-up, approach and Tables 1 to 3, and from the four CVE records. Red cards are not stated by the report.

Findings were not passed straight on. The report says that, for a limited number of codebases, ENISA reviewed the AI findings, "developed controlled proofs of concept" and tested them manually to judge impact and reproducibility, removed duplicates, and then reported through coordinated vulnerability disclosure: for codebases maintained by Union entities through CERT-EU, which acts as the coordinator, and otherwise under each project's own policy. Maintainers were asked to say whether each finding was a true positive and to validate its severity. All findings were shared with code owners in August 2026 (report).

From 220 to four, on one axis

The diagram puts the four counts on one zero based axis. The report prints 220, 88 and 4 in its Table 3; 84 is this site's subtraction from Figure 6. The dashed part of the second bar is the 132 with no answer, and it is the part the 95% says nothing about.

Drawn from the report's Table 3, summary and Figures 3 and 6. The scale comes from this site's own scale code. 84 and the percentages are derived.

The counts, where each comes from and the arithmetic. Figures not printed by the report are labelled derived.

CountValueSource and arithmetic
Codebases analysed27 (25 already open source, 2 soon to be)Report summary and Table 3
AI-identified findings220Report summary and Table 3
Assessed by code owners at publication88, 40%Report summary, Figure 3, Table 3. Derived: 88 / 220 = 40.0%
Without an answer at publication132, 60%Derived: 220 - 88 = 132. Figure 3 prints 60% unverified
Marked false by code owners4Figure 6 label. Figure 4 prints 5% false, derived 4 / 88 = 4.5%
Confirmed true positives84, 95.5% of 88Derived: 88 - 4 = 84; 84 / 88 = 95.5%. The report prints 95%
CVE ids assigned4Report Table 3 and the report's text (to one codebase). The four records are in the next sections
AI severity mix of the 2204% critical, 44% high, 34% medium, 18% lowReport Figure 2, as the AI tools rated them
AI severity mix of the 883, 39, 30, 16 findings: 3.4%, 44.3%, 34.1%, 18.2%Report Figure 6 counts. Derived shares. They match the 220's mix within a point, so the 88 do not look skewed on the AI's own ratings; that says nothing about whether they are true

Spread over the codebases, 220 findings is 8.1 per codebase and 88 assessed is 3.3 (derived: 220 / 27 and 88 / 27). These are averages: the report prints no count per codebase, so the spread, the largest batch and the number of codebases with a confirmed finding are not stated. The report says assessing AI findings "requires dedicated expert time from code owners" and that the volume "can place a significant review burden on code owners". It prints no hours.

On timing, the report prints "August 2026" for when all findings were shared and "the time of publication" for the cut-off of the 88; its own date is "November 2026". It prints no day for any of them, so this site does not work out a number of weeks or months. It says the assessments were still ongoing. The four CVE records are dated 12 August 2026, so at least one finding had reached a CVE id by then.

What it cost, in the report's indicative figures

The report says the source code reviews were conducted in August 2026 using several models, so the actual cost varies, and that the total cost of all the scans is "roughly estimated" at around EUR 6 648.90, with an average code audit cost of about EUR 132.27 per 10 000 lines of code. It adds that the figures are indicative only and that access to the models was free of charge under the providers' programmes. Its Table 3 prints 611,486,042 as the total tokens consumed. Every euro figure below is therefore this site's division of an indicative estimate, not a price anyone paid.

Cost and token measures worked out from the report's Table 3 and cost estimate. All derived and indicative; the report's access to the models was free.

Measure (derived)ValueArithmetic
Per AI findingEUR 30.226,648.90 / 220
Per finding assessed by code ownersEUR 75.566,648.90 / 88
Per confirmed findingEUR 79.156,648.90 / 84
Per codebaseEUR 246.266,648.90 / 27
Per CVE idEUR 1,662.226,648.90 / 4
Tokens per AI finding2.78 million611,486,042 / 220
Tokens per confirmed finding7.28 million611,486,042 / 84
Tokens per codebase22.6 million611,486,042 / 27
Blended price per million tokensEUR 10.876,648.90 / 611.486042

The token column. Table 3's header reads "Total tokens consumed (M)" and its cell prints 611,486,042. If the (M) meant millions the cell would be 611 trillion tokens, which cannot sit beside a cost estimate of a few thousand euros. This site reads the cell as a raw count of about 611 million tokens and says that is its reading; the report does not explain the header.

Two denominators. The report's EUR 132.27 is per 10 000 lines, this site's EUR 246.26 is per codebase, and neither is the other. They are also not a ratio of totals: dividing 6,648.90 by 132.27 and multiplying by 10 000 would imply only about 502,676 lines in all (derived), while Figure 8 has points at about a million lines and above. The report does not say how the 132.27 was averaged.

Reading Figure 8. The report prints no table behind its two scatter plots of cost and run time against lines of code, both on log axes. This site counted the points on the rendered page: 27 in each panel, 13 for general purpose agents with skills or plug-ins and 14 for Vigilo, the same number as the codebases, though the report does not say a point is a codebase. Read against the axes the cost points run from about EUR 17 to about EUR 329 and add up to roughly EUR 2,774, about 42% of 6,648.90 (derived, rough: positions read off a log axis, accurate to a few percent of the axis). The points sum to roughly 5.1 million lines. On those readings the total cost over the total lines is about EUR 5.5 per 10 000 lines, while the mean of the per-point ratios is about EUR 102, much nearer the report's 132.27; that fits an average of per-codebase ratios, which small codebases dominate. The report does not say whether Figure 8 shows every run, which is a further reason to treat all of this as rough.

Which tool costs more. The report's text says the general purpose agents were usually faster and cheaper than Vigilo, and Figure 8 shows the Vigilo points sitting higher in both panels. This site's rough mean is about EUR 84 a point for the 13 general purpose points and about EUR 120 for the 14 Vigilo points (derived, rough). The two sets are different codebases of different sizes, so that is not a like for like price. The report gives no cost per finding or per confirmed finding for either tool type: not stated.

The NATO number. OpenAI's page says NATO has used "up to 300 billion raw input and output tokens per month" in its work with OpenAI. Beside 611,486,042 tokens that is about 490 times as many in a month (derived: 300,000,000,000 / 611,486,042 = 490.6). It is a vendor's figure for a customer, an upper bound ("up to"), for different work and a different period, and it is not a benchmark for anything in the report. The page also quotes NATO's communications and information agency saying that as adoption scales it will seek "a deeper understanding of token consumption" and will "evaluate competing solutions": the customer says it does not yet have that understanding.

What the cost leaves out: the report says setting up a secure environment "takes considerable time" and needs continuous monitoring, and that code owners need time and resources to assess findings. No hours are printed for ENISA's staff, for CERT-EU's, or for the code owners. The seventh step of the report's annex is to monitor token usage and operational costs; this site's suggestion further down is to record them per confirmed finding.

The AI's severity against the code owners'

Figure 6 is the report's one picture of what review changed. Its left side gives the AI tools' severity for the 88 assessed findings: 3 critical, 39 high, 30 medium, 16 low. Its right side gives the code owners' rating: none critical, 22 high, 29 medium, 24 low, 9 informational and 4 marked false. Both sides add to 88. That is 42 rated critical or high by the AI against 22 rated high by the owners (derived: 3 + 39 = 42; 22 / 42 = 52%). The report's own word for this is "a slight tendency" to overrate.

Redrawn to scale from counts printed on the report's Figures 6 and 7, read from the rendered pages. The scales come from this site's own scale code. Sums and shares marked derived are this site's arithmetic.

Only three of the figure's flows carry a printed count for findings that kept their severity: 21 (high to high), 22 (medium to medium) and 14 (low to low). They add to 57 (derived), which is 64.8% of the 88, or 67.9% of the 84 confirmed. So about 27 of the 84 confirmed findings changed severity (derived: 84 - 57). That is consistent with the two tool types' kept shares in Figure 7, 60% and 74%. The other flows are drawn but unlabelled, and this site does not count them.

Figure 7 gives the shares for confirmed findings by tool type. Code owners kept the severity Vigilo assigned in 74% of cases (23% downgraded, 3% upgraded), against 60% (36% and 4%) for general purpose coding agents with a security skill or plug-in. The report reads this as the extra steps of a specialised harness, such as validating findings with an adversarial agent, producing ratings closer to the owners' own, and adds that in both cases most changes were downgrades. The report prints no counts per tool type, no interval and no note of whether the same codebases were run by both, and Vigilo is the harness CERT-EU, a co-author, developed. Those are the report's results from 27 codebases, not a general ranking.

The text against the figures. The report says the reviewed findings were rated 26% high, 34% medium, 29% low and 11% informational. Those shares fit 22, 29, 24 and 9 out of the 84 confirmed (26.2%, 34.5%, 28.6%, 10.7%; derived) and do not fit 88, where the same counts give 25%, 33%, 27% and 10%. The medium share rounds to 35% on 84, and the report prints 34%. The difference is small and is listed with the others below.

Why the severity moves. The report's second observation is that the severity of a finding can rarely be determined from source code alone, because a model needs to know what is reachable and with which privileges, which components are trusted and what external controls exist. It also says the 5% classed as false positives or not applicable were so "mostly because the AI tools lacked information about the runtime environment, which was deliberately withheld from them". The experiment withheld deployment context by design, so part of what it measures is the cost of that choice. Readers who give a tool deployment context should not assume the same shift.

The four CVE records: what they print

The report prints no CVE ids and names no product. Its text says the code owners' feedback "also led to four new CVE IDs being assigned to one of the scanned codebases as a result of the experiment". This site found the four ids through OpenAI's page, which links them, and read the records at CVE.org's public API (CVE-2026-73374, CVE-2026-73405, CVE-2026-73431, CVE-2026-73432). That is a public source and a secondary route: the report does not say these are the four, and this site matched them through OpenAI's page, which links the four records beside its account of the report. The records credit an ENISA team as reporter and the report says four CVE ids were assigned, so the match is reasoned and not printed by the report.

All four records were assigned by CIRCL, were published on 12 August 2026, name one product, vulnerability-lookup, and list every version up to and including 5.5.1 as affected. Each describes one flaw in a sentence, which this site quotes by the record's own phrase and summarises in one line, without any step to reproduce it.

The four CVE records as printed: id, the record's title, the CWE and the CVSS 4.0 base score the record carries. Read on 17 November 2026.

CVE idTitle as printedCWECVSS 4.0 score
CVE-2026-73374Stored Cross-Site Scripting (XSS) via Unescaped CNA Reference Tags in vulnerability-lookupCWE-796.1 Medium
CVE-2026-73405Authorization Bypass in SSE Pub/Sub Allows Unconfirmed Accounts to Access Stream Events in vulnerability-lookupCWE-8625.3 Medium
CVE-2026-73431Reusable Account Activation and Recovery Tokens Allow Repeated Account Takeover in vulnerability-lookupCWE-2948.8 High
CVE-2026-73432Stored Server-Side Request Forgery in Remote-Instance Synchronization Allows Access to Internal Services in vulnerability-lookupCWE-9185.1 Medium

In one line each, from the records' descriptions. CVE-2026-73374 is a "stored cross-site scripting (XSS) vulnerability" in how reference tags are shown, fixed by escaping each tag. CVE-2026-73405 is an "authorization bypass vulnerability": a live-event subscription endpoint did not check that an account was active and confirmed. CVE-2026-73431 concerns activation and recovery links that stayed valid after use for their whole validity period. CVE-2026-73432 is a "server-side request forgery (SSRF) vulnerability" that needs an administrator permission. Each record states a precondition: record editing rights, a self-registered account, a captured link or an administrator permission.

Credits, and what they establish. Every record credits the same tool, "openai/gpt-5.5-cyber-preview (reasoning level: high)", and the same reporter organisation, "EUVD @ ENISA Team". That is the only place in the saved pages where a model is tied to a finding. The report lists Anthropic's Mythos 5 among the models and says 220 findings came from the models used; it does not split them, so a record crediting one OpenAI model does not show what any other model found. The records also credit a remediation reviewer and a coordinator by name; this site prints neither.

Four ids, one product. Four ids from 84 confirmed findings is 4.8% (derived: 4 / 84), or 1.8% of the 220. The report says all four were assigned to one scanned codebase, and the records name one product, so the four ids say nothing about the other 26 codebases. The report does not say how many of the 27 codebases had a confirmed finding. Each id carries a CVSS 4.0 score of 5.1 to 8.8; those scores and the code owners' high, medium and low in Figure 6 are different scales, and the report does not say which of the 88 findings became the four ids.

Fixes. OpenAI's page says "fixes published for all four". Each record links one commit marked as a patch, and CVE-2026-73405 and CVE-2026-73431 link the same commit, so the four records link three distinct commits. No record prints a fixed release number, and each lists versions up to and including 5.5.1 as affected. This site did not read the commits or the project's release notes.

Exploitation. The report does not say whether any finding has been exploited. Each record carries an entry from CISA-ADP (an enrichment entry titled CISA ADP Vulnrichment, not the assigner) with a timestamp of 12 August 2026 that lists Exploitation as none, Automatable as no and Technical Impact as partial, or total for CVE-2026-73431. That is a statement as of that timestamp. This site did not check later status and does not claim the flaws were or were not exploited.

Text
CVE-2026-73374  CWE-79  CVSS 4.0 base score 6.1 MEDIUM
  CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:A/VC:L/VI:L/VA:N/SC:H/SI:N/SA:N
CVE-2026-73405  CWE-862  CVSS 4.0 base score 5.3 MEDIUM
  CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:L/VI:L/VA:N/SC:N/SI:N/SA:N
CVE-2026-73431  CWE-294  CVSS 4.0 base score 8.8 HIGH
  CVSS:4.0/AV:N/AC:H/AT:N/PR:L/UI:N/VC:H/VI:H/VA:N/SC:H/SI:H/SA:N
CVE-2026-73432  CWE-918  CVSS 4.0 base score 5.1 MEDIUM
  CVSS:4.0/AV:N/AC:L/AT:N/PR:H/UI:N/VC:L/VI:L/VA:N/SC:L/SI:L/SA:N
The CWE, CVSS 4.0 base score and vector string from each of the four CVE records, as read on 17 November 2026. The vectors are the records' own scores, assigned by CIRCL as the record's source.
Text
affected product:   vulnerability-lookup
affected versions:  lessThanOrEqual 5.5.1 (status affected, version type semver)
assigner:           CIRCL
published:          2026-08-12 (all four records)
credit, tool:       openai/gpt-5.5-cyber-preview (reasoning level: high)
credit, reporter:   EUVD @ ENISA Team
credit, other:      a remediation reviewer and a coordinator (named in the records; not printed here)
The affected range, assigner, publication date and credit lines as the four CVE records print them (identical across the four), read on 17 November 2026. The two personal credits are left out by this site.
Text
CODEBASES  AI-IDENTIFIED FINDINGS  ACKNOWLEDGED BY CODE OWNERS  CVE ID ASSIGNMENTS  TOTAL TOKENS CONSUMED (M)
27         220 (for expert triage)  88                          4                  611,486,042
The report's Table 3 row as printed (the column headings joined from a table whose headings wrap across lines), read on 17 November 2026. The header says (M) and the cell is a raw looking count: see the cost section.

The report prints no file hashes, indicators, payloads or live infrastructure, and this brief carries none. The blocks above are the records' fields and the report's table, and nothing in them reproduces a flaw.

Where the pages differ

Eleven points where the report, the records and OpenAI's page differ, leave a gap or use a different basis. This site resolves none of them.

Differences and gaps between the report, the CVE records and OpenAI's page, as read on 17 November 2026.

PointOne page printsThe other page prints, or what is missing
Token unitReport Table 3 header: total tokens consumed (M)The same table's cell: 611,486,042. Read here as a raw count
95%Report summary and Figure 4: 95% true, 5% false of the 88Figure 6 marks 4 of 88 false, 95.5% (derived); the text rounds
Severity sharesReport text: 26, 34, 29, 11 percent after reviewFigure 6 counts 22, 29, 24, 9 plus 4 false; the shares fit 84, not 88, and medium rounds to 35 on 84
Which modelOpenAI: ENISA used its models to identify four vulnerabilities; the report is described as covering OpenAI models through DaybreakReport Table 1 also lists Anthropic's Mythos 5 and does not say which model found which finding. The four records credit one OpenAI model
FixesOpenAI: fixes published for all fourThe four records link three distinct commits and no fixed release number. Not read by this site
Where the four ids sitReport: one of the scanned codebases. OpenAI: software used across EU institutionsThe records name vulnerability-lookup, assigner CIRCL. Whether a Union entity maintains it is not printed in the saved pages
WhenReport: reviews running since June 2026; source code reviews conducted in August 2026; findings shared in August 2026No day printed for any of them. The records are dated 12 August 2026
Report dateReport: November 2026OpenAI's page, dated 17 November 2026, calls it a November joint report. The day of publication is not printed
Who wrote whatThe report: ENISA and CERT-EU wrote it and tested CERT-EU's harness and ENISA's skillsThe comparison of Vigilo with general purpose tools is by its authors, on 27 codebases
NATO's tokensOpenAI: NATO has used up to 300 billion raw tokens a month in its work with OpenAIThe report prints no NATO figure. NCIA's own quote speaks of building an understanding of token consumption
Document draftingReport header: AI tools were used in a limited capacity to support editorial refinementOutputs reviewed by subject matter experts, as the report says. Reported here once, not as a criticism

What the report states, and what it does not

What is stated and not stated about the experiment, from the report, the records and OpenAI's page.

QuestionStatedNot stated
How many findings are true95% of the 88 assessed; 84 of 88 by Figure 6 (derived)How many of the other 132 are true; how the 88 were chosen
Which model found which findingModels, harnesses and skills used in various combinationsAny count per model or per harness, or which found the four that became CVE ids beyond the records' tool credit
Which productsFour CVE ids in one codebase; the records name vulnerability-lookupAny other product; how many of the 27 codebases had a confirmed finding
SeverityAI against owners for the 88; shares by tool typeSeverity of the 132; owners' CVSS; counts per tool type
CostAbout EUR 6,648.90 and EUR 132.27 per 10 000 lines, indicative; free accessHours for code owners, ENISA or CERT-EU; set-up effort; cost per finding by tool type
Cost by tool typeFigure 8 scatter plots of cost and run time against lines of codePer finding or per confirmed finding for Vigilo or the general purpose agents
ExploitationCISA-ADP entries of 12 August 2026 list Exploitation as noneAny later status; the report says nothing on exploitation
FixesThe records link commits; OpenAI says fixes were publishedA fixed release number; whether the fixes reached users
DatesSince June 2026; shared in August 2026; report November 2026Any day other than 12 August 2026; assessment dates; the report's day

What to do, in order

What ENISA and CERT-EU recommend (the report's observations, lessons and annex). Run AI code review in a secure, isolated environment with network segmentation, clearly defined trust boundaries, access control and monitoring. Have qualified human experts review the results and their severity. Deduplicate and, as far as possible, verify AI output before reporting it, and say when full verification needs deployment context only the owner has. Report through coordinated vulnerability disclosure: for Union entities through CERT-EU, for community maintained projects under each project's own policy. The annex adds a seven step cycle: select or develop skills and harnesses, run reputable models, deduplicate and triage, develop proofs of concept in a controlled environment, report confirmed issues, generate remediation guidance for owners to review, and monitor token usage and cost.

Take this with you

This site's suggestions, in order

  • Maintainers: before you accept a batch of AI findings, decide who triages them and how many hours a week that person has. The report says owners need time and resources, and prints no hours; the 8.1 findings per codebase average is derived, not a workload.
  • Maintainers: ask the reporter which model, which harness and what deployment context were withheld. The report shows severity moving down when context is missing, so a rating without that note is weaker evidence.
  • Maintainers: ask for a reproducer or a stated reason there is none, and keep a written count of reports received, confirmed, rejected and fixed, so that your true positive rate is yours and not a vendor's or an agency's.
  • Security teams running AI code review: record tokens and euros per confirmed finding, not per finding, and the hours your reviewers spend. The report prints per finding and per codebase figures only by this site's division.
  • Security teams: run the tools in an isolated environment as the report says, and compare a general purpose agent baseline against a specialised harness on your own code before choosing. The report's comparison is its authors' tools on 27 codebases.
  • Buyers: do not treat a vendor's customer token count, such as the NATO figure on OpenAI's page, as a benchmark. It is an upper bound, for another job, from the seller.
  • Readers of the headline: 95% is of the 88 answered findings, not of 220, and the report says AI severity ratings need human review before they are relied on.
  • Operators of Vulnerability-Lookup: the records list versions up to and including 5.5.1 as affected. Check the project's release notes for a fixed version and update; this site did not read them.

The question the report leaves open

The report's lessons section says finding got faster and validating, prioritising and fixing "cannot currently be accelerated to the same extent", and that the shortage of experts able to verify AI findings "will become an increasingly serious constraint". Its own count shows the gap: 132 findings without an answer at publication, and no hours printed for the 88 that had one. If an agency's AI scan hands you 220 findings, how many hours do you have to answer them, and who counts how many were true?

Sources

  1. PrimaryJoint advisory on the use of frontier AI models for code review, 15 pages, November 2026: the set-up, Tables 1 to 3, Figures 1 to 8, the observations, the lessons and the annex. Read in full by this site on 17 November 2026ENISA and CERT-EUaccessed 2026-11-17
  2. PrimaryCVE-2026-73374, published 12 August 2026: vulnerability-lookup up to 5.5.1, CWE-79, CVSS 4.0 of 6.1, credits and the linked fix commitCVE Program (assigned by CIRCL)accessed 2026-11-17
  3. PrimaryCVE-2026-73405, published 12 August 2026: vulnerability-lookup up to 5.5.1, CWE-862, CVSS 4.0 of 5.3, credits and the linked fix commitCVE Program (assigned by CIRCL)accessed 2026-11-17
  4. PrimaryCVE-2026-73431, published 12 August 2026: vulnerability-lookup up to 5.5.1, CWE-294, CVSS 4.0 of 8.8, credits and the linked fix commitCVE Program (assigned by CIRCL)accessed 2026-11-17
  5. PrimaryCVE-2026-73432, published 12 August 2026: vulnerability-lookup up to 5.5.1, CWE-918, CVSS 4.0 of 5.1, credits and the linked fix commitCVE Program (assigned by CIRCL)accessed 2026-11-17
  6. Reported byStrengthening NATO and allied cyber defenses with AI, page dated 17 November 2026: cites the report and the four CVE ids, says fixes were published for all four, and gives NATO's token figure. A vendor statement and the route by which this site matched the four ids. SecondaryOpenAIaccessed 2026-11-17

Share this briefing

Know someone who owns this problem? Send it to them.

Related briefings

The briefing, in your inbox

Practitioner analysis of cyber and AI security news. No vendor noise.

How often

Every new briefing in one email, at 7am, or at 7am, 12:30pm and 6pm. Nothing is sent when nothing is new. Unsubscribe any time.