The Receipts Room
Picture the one person who runs IT for a 30-person manufacturer somewhere in Europe. In July, Europe’s cybersecurity agency, ENISA, warned people like her that the time between a software flaw being discovered and attackers using it could soon shrink from months to hours, or even minutes. Her biggest weakness, it said, would no longer be the technology. It would be the meeting where someone has to approve the fix.
ENISA also said more help was on the way. She has been waiting for it. So have we.
Some background. Our look at ENISA’s 2026 threat report promised a separate look at its paper on the most capable AI models. We delivered that in From Years to Minutes, which ended with a promise of our own: ENISA said it would refine its advice, and we would report what changed. This is that report, as of 19 November 2026.
Our proposition: ENISA is keeping its promise slowly, and mostly through side doors. Its first advice under the EU’s new plan arrived six weeks after the deadline, and it helps experts find flaws faster. The help it promised to the organizations stuck waiting for a sign-off has still not arrived.
A promise made in July, with a deadline attached
ENISA’s July paper, ENISA’s view on Cybersecurity in the Frontier AI Era, was modest about itself. It described its advice as a first set of recommendations, not a complete checklist, and committed to refining and expanding them together with EU member states and the EU’s own institutions, in step with a European Commission action plan that was still to come (European Union Agency for Cybersecurity [ENISA], 2026a, p. 5).
That action plan was not really still to come. The Commission presented its EU Action Plan on Cybersecurity and Artificial Intelligence on the very same day, 7 July 2026 (European Commission, 2026a). And it gave ENISA a job with a date on it. Under what the plan calls Key Action 4, ENISA was to start issuing guidance, recommendations, advisories and best practices, both on protecting against AI-powered attacks and on using AI safely in security work, from the third quarter of 2026 (ENISA & CERT-EU, 2026f, p. 5; European Commission, 2026b). The plan also commits the Commission, ENISA, member states and industry to bring vulnerability management up to speed for the AI era, and lists a pilot campaign to speed up patching of the most critical open-source software for the fourth quarter of 2026 (European Commission, 2026b).
The third quarter ended on 30 September.
Nine publications later, only one is about AI
On 19 November we went through ENISA’s publications list. Since 7 July it has published nine documents: a maturity model for small and medium-sized businesses (13 July), procurement guidelines for hospitals (22 July), a Secure by Design and Default Playbook (30 July), the Threat Landscape 2026 (22 September), a technical advisory on software bills of materials (22 October), a technical advisory on secure update mechanisms (4 November), the 2026 Report on the State of Cybersecurity in the Union (10 November), a joint advisory on using frontier AI models to review code (12 November) and a document on accrediting certification bodies under the Cyber Resilience Act (16 November). Only one of them, the 12 November advisory, is filed under ENISA’s artificial intelligence topic (ENISA, n.d.-a, n.d.-b).
Before we get to that one, it is worth noticing that AI runs through four of the others, even where the title does not say so.
The flood of fixes, named. ENISA’s technical advisory on secure update mechanisms (4 November) expects AI-assisted tools for finding vulnerabilities, and it names Anthropic’s Claude Mythos among them, to increase the number of flaws that manufacturers must assess, rank and patch. That, the advisory argues, makes it even more important that updates reach customers securely, quickly and reliably (ENISA, 2026b, p. 4). It is written for the makers of connected products covered by the EU’s Cyber Resilience Act, not for the people who buy and run them.
A new pipe for reporting flaws. On 11 September ENISA switched on the Single Reporting Platform, where manufacturers must now report actively exploited vulnerabilities and severe incidents under that same law (ENISA, 2026c). The July paper had pointed to this platform as one way to put some of its recommendations into practice (ENISA, 2026a, p. 13).
A bigger role in numbering flaws. On 6 August ENISA announced a larger role in the CVE Program, the international system that gives each known vulnerability an official identifier. It now oversees 20 organizations that hand out those identifiers for European entities, and it named frontier AI’s effect on how flaws are found and exploited as a reason to build up that capacity (ENISA, 2026d).
A diagnosis, but no prescription. ENISA’s 2026 Report on the State of Cybersecurity in the Union describes frontier AI as turning the discovery of vulnerabilities into something close to an industrial process (ENISA, 2026g, p. 8), and says it is pushing authorities and industry worldwide to rethink how vulnerabilities are handled from start to finish (p. 50). The press release that came with it describes AI as boosting both attackers and defenders (ENISA, 2026e). Yet what the report asks of organizations is the familiar list: build securely from the start, have a process for disclosing flaws, and patch promptly (p. 50). It does not refine the July recommendations.
The advice that finally arrived, six weeks late
On 12 November ENISA and CERT-EU, the cybersecurity service for the EU’s own institutions, published a Joint Advisory on the Use of Frontier AI Models for Code Review. It states that the work was done under Key Action 4 of the EU plan (ENISA & CERT-EU, 2026f, p. 5). That makes it the first publication that can be tied to the third-quarter task, and it arrived six weeks after the quarter ended.
It is an experiment, and an honest one. ENISA ran frontier AI models from OpenAI, including its cybersecurity versions of GPT-5.5 and GPT-5.6, and from Anthropic, its Mythos 5 model, over the source code of 27 open-source software projects that EU bodies maintain. No live systems were touched. The models flagged 220 possible problems. By the time the advisory was published, the people responsible for the code had checked 88 of them and confirmed 95 percent as real, and four new official vulnerability identifiers had been issued (ENISA & CERT-EU, 2026f, pp. 5–7).
Its central lesson is the one the July paper warned about, now seen from the inside. The AI made finding flaws much faster, but it could not replace human judgement, and the sheer number of findings landed as extra work on the people who maintain the code. Checking, ranking and fixing the problems, the advisory says, cannot yet be sped up to anything like the same degree (ENISA & CERT-EU, 2026f, pp. 5, 13). Its advice for anyone trying the same thing is sensible: work in a secure, isolated test environment, have qualified people review every result and how serious it really is, and follow an established process for disclosing what you find (p. 5).
That is useful, and it is real work. But it is advice for the people who find flaws, not for the ones who have to approve the fix.
The gap nobody has closed: getting the fix approved
Three things are still missing from the public record. First, there is no refined or expanded version of the July recommendations. Second, the third-quarter task has so far produced one publication, late, and it covers finding flaws rather than defending against AI-powered attacks more broadly. Third, nothing new speaks to the idea that made the July paper worth reading: what ENISA called the Authority Gap, meaning organizations that cannot approve a fix as fast as an attacker can use the flaw (ENISA, 2026a, p. 8). The new advisory describes that bottleneck from the inside. It does not say what to do about it.
The July paper also floated EU-wide benchmarks for testing the cyber abilities of advanced AI models (ENISA, 2026a, p. 14). Nothing has been published on those either.
Big companies can cope. Small ones are left waiting
Large companies with their own security teams do not need ENISA to tell them how to speed up their change approvals. The organizations that do are the ones ENISA itself singled out in July: smaller ones, short on money, security focus and skills (ENISA, 2026a, p. 8). What has been published so far helps manufacturers deliver safer updates and helps experts find flaws faster. It does much less for the people who have to approve and install the fixes. In our reading, that is where the delay costs the most.
Is four months really too long? The case for patience
There is a fair case for ENISA. Four months is not long for an EU agency that promised to work closely with member states and the EU’s institutions; consultation is slow because it is meant to be. Some of its work goes straight to national authorities and incident-response networks and is never published, so guidance may exist that the public simply cannot see. And a working reporting platform, a clear advisory on secure updates and a careful experiment with frontier AI are arguably worth more than another paper. They are the practical plumbing the July paper asked for.
All true. But ENISA set the bar itself, in public, and the Commission put a date on it.
Our verdict: promise partly kept, and late
The practical commitments are real and arrived on time. The first advice under the EU plan arrived on 12 November, six weeks past its quarter, and it is solid work on finding flaws. The specific promise, refined and expanded recommendations on frontier AI, had still not appeared in public on 19 November 2026. There is an irony worth naming, gently: the agency that diagnosed the Authority Gap is living a version of it. Its consultations run at committee speed. The threat it described does not.
Questions ENISA has not yet answered in public
- Whether ENISA considers its third-quarter task done with the 12 November advisory, or whether more is coming. We have not asked it; this report rests on what ENISA and the Commission have published.
- What the code owners concluded about the 132 findings they had not yet checked.
- Whether guidance has gone to national authorities without being published.
- Whether the fourth-quarter items, including the open-source patching pilot, arrive by 31 December.
What you can do before the guidance arrives
You do not need to wait for Brussels to close your own Authority Gap. Write down, this week, which kinds of security update can go out without a meeting, which need one named person to approve them, and which still need the full board. Then ask the suppliers of your most important software how they sign and deliver their updates, and how quickly critical fixes reach you. ENISA’s 4 November advisory is written for those suppliers; its questions work just as well coming from you.
The new question is simple. When the year ends on 31 December, will Europe have given defenders anything that works at the speed it described in July? We will check in January.
Get the follow-up. The free Mercury Brief tells you when we report back on promises like this one. Get the free Mercury Brief.
References
European Commission. (2026a, July 7). Commission presents EU Action Plan on Cybersecurity and Artificial Intelligence [News]. https://digital-strategy.ec.europa.eu/en/news/commission-presents-eu-action-plan-cybersecurity-and-artificial-intelligence
European Commission. (2026b). Action plan on cybersecurity and artificial intelligence [Factsheet]. Publications Office of the European Union. https://doi.org/10.2759/9909000. https://ec.europa.eu/commission/presscorner/api/files/attachment/882653/Factsheet%20-%20Action%20Plan%20on%20Cybersecurity%20and%20AI.pdf
European Union Agency for Cybersecurity. (2026a, July 7). ENISA’s view on cybersecurity in the frontier AI era. ENISA. https://www.enisa.europa.eu/sites/default/files/2026-07/ENISA%20view%20on%20cybersecurity%20in%20the%20frontier%20AI%20era_en_0.pdf
European Union Agency for Cybersecurity. (2026b, November 4). ENISA technical advisory on secure update mechanisms (Version 1.0). ENISA. https://www.enisa.europa.eu/sites/default/files/2026-11/ENISA%20Technical%20Advisory%20-%20Update%20Mechanisms%20-%20v1.pdf
European Union Agency for Cybersecurity. (2026c, September 11). The CRA Single Reporting Platform is launched [Press release]. https://www.enisa.europa.eu/news/the-cra-single-reporting-platform-is-launched
European Union Agency for Cybersecurity. (2026d, August 6). ENISA scales up its role in the CVE Program [Press release]. https://www.enisa.europa.eu/news/enisa-scales-up-its-role-in-the-cve-program
European Union Agency for Cybersecurity. (2026e, November 10). ENISA reveals the State of Cybersecurity in the Union [Press release]. https://www.enisa.europa.eu/news/enisa-reveals-the-state-of-cybersecurity-in-the-union
European Union Agency for Cybersecurity & CERT-EU. (2026f, November). Joint advisory on the use of frontier AI models for code review. ENISA. https://doi.org/10.2824/7940736. https://www.enisa.europa.eu/publications/joint-advisory-on-the-use-of-frontier-ai-models-for-code-review
European Union Agency for Cybersecurity. (2026g, November). 2026 report on the state of cybersecurity in the Union. ENISA. https://www.enisa.europa.eu/publications/2026-report-on-the-state-of-cybersecurity-in-the-union
European Union Agency for Cybersecurity. (n.d.-a). Publications. Retrieved November 19, 2026, from https://www.enisa.europa.eu/publications
European Union Agency for Cybersecurity. (n.d.-b). Artificial intelligence and next gen technologies. Retrieved November 19, 2026, from https://www.enisa.europa.eu/topics/artificial-intelligence-and-next-gen-technologies



Leave a Reply
You must be logged in to post a comment.