The Control File
You run a small business. A security update for your office firewall lands on a Tuesday. Your IT contractor usually installs updates at the end of the month, after checking nothing will break. That routine has served you for years. (That business is an illustration. The question it raises is not.)
Europe’s cybersecurity agency now says that routine may be the most dangerous thing in your building.
For most of the history of computing, there was a grace period. Someone would find a flaw in a piece of software. Weeks or months might pass before anyone turned it into a working attack, and in that time the maker could ship a fix and most people could install it. Security teams built their routines around that gap: test the patch, wait for the monthly maintenance window, get approval from a change board, roll it out.
The European Union Agency for Cybersecurity, ENISA, says that grace period is disappearing. In a short paper published on 7 July 2026, ENISA’s view on Cybersecurity in the Frontier AI Era, it writes that the time between a vulnerability being discovered and being exploited “has gone from years, to months, and now (potentially) to hours or even minutes” (European Union Agency for Cybersecurity [ENISA], 2026, p. 7). The reason is frontier AI: the most capable AI models, which can find flaws and work out how to use them far faster than people can.
Here is my proposition: in the AI era, the weakest link in your security is no longer the software. It is the time it takes someone to say yes to the fix. Attackers benefit from every hour of that wait. The people who pay are the ones with the least slack: small firms, stretched IT teams, and anyone with a device that no longer gets updates.
What ENISA actually published
The paper is 16 pages long, written for national authorities, EU policymakers, defenders and service providers. ENISA describes it as an “initial set of recommendations”, not an all-inclusive checklist, drawn from engagement with EU cyber crisis and incident response networks, industry and the open-source community (ENISA, 2026, pp. 2, 5). It is an assessment and a set of recommendations, not a report on incidents.
ENISA expects the capability to spread. It says openly available models may reach a similar level within 9 to 12 months (p. 5). Among the concerns stakeholders raised with it: attackers may have exploits before fixes exist, what the paper calls “negative time-to-exploit”; older products that have reached or will soon reach end of support will be harder to protect; more frequent patches may mean more service disruptions; and small and medium-sized businesses may need extra help (p. 5).
From years to minutes: the receipts
ENISA points to industry research suggesting attackers can weaponise new vulnerabilities within 15 minutes of disclosure, and that the median time from first breaking in to stealing data has fallen to 72 minutes (ENISA, 2026, p. 7, citing industry reporting). It quotes the EU institutions’ own computer emergency response team, CERT-EU, on why the newest models matter: not just the volume of flaws they find, but their ability to chain findings together. “This is what turns a list of individual flaws into a working attack” (p. 8).
The result is a flood. ENISA relays one industry representative’s account: their organisation used to receive a few hundred possible vulnerability reports a year, of which about a hundred would normally earn an official identifier. It went from about 80 of those identifiers in the first quarter of 2025 to close to 500 in the first quarter of 2026, and to “about 500 per day” once it used frontier AI tools (pp. 8–9).
Open-source projects have felt it too. By early 2026, ENISA says, AI-generated reports had begun to overwhelm parts of the open-source disclosure pipeline, one bug bounty platform paused submissions, and the widely used curl project shut its coordinated disclosure programme in January 2026 because maintainers could not absorb the volume (p. 10). ENISA adds that, according to its recent discussions with industry, the quality of AI-generated reports has improved markedly, so much of the flood is now genuine signal (p. 10). That is the uncomfortable part: it is not mostly noise.
Key idea: The window of exposure
The window of exposure is the time between a flaw in software becoming known to someone and the fix actually being installed on your device or system. While that window is open, anyone who knows about the flaw can use it. Security has always been a race to close the window faster than attackers can use it. If AI shrinks the attacker’s side of the race to minutes, then the slowest part of the defence, often waiting for someone to approve or install an update, becomes the main risk.
The meeting is now the vulnerability
This is the idea in the paper most worth borrowing. ENISA argues that, in the new environment, “the primary latency is no longer technical, but procedural.” It calls this the “Authority Gap”: the inability of human change advisory boards, the committees that approve changes to important systems, to authorise a fix in the few minutes now available. “Plainly,” the paper says, “it means organisations are too slow to authorise a meaningful antidote” (ENISA, 2026, p. 8).
Who gets squeezed hardest? ENISA is blunt. Larger organisations will probably adapt faster; smaller ones “will face challenges due to a lack of resources, security focus, and the necessary skills” (p. 8). The firm with a monthly contractor visit is exactly the firm the paper is worried about.
The complication: just automate it?
The obvious answer is to let machines apply fixes automatically. ENISA does not pretend that is simple. Critical service providers “may be forced” to patch autonomously because organisational approval is too slow, but automated updates can break systems and cause downtime, and checking AI-written patches may itself become a bottleneck (p. 11). Anyone who has had an update take down a till system on a Saturday knows why change boards exist.
And notice what ENISA does not say. It does not say remove people. Elsewhere it insists that solutions keep “the human in the middle”, calls for “human-gated triage” in AI-enabled security tools, and asks defenders to keep “human review firmly in the loop” even as response times fall to minutes (pp. 12, 14, 15). The tension is real: approvals protect against bad changes, and slow approvals now carry their own risk.
The verdict: decide before the emergency
My read of the evidence: the answer is neither “remove the human” nor “keep the monthly meeting.” It is to make the hard decisions before the emergency. If an organisation has already agreed which kinds of fix can go straight out, which need one named person, and which still need a full board, it does not have to convene a committee at 2 a.m. That is the logic of approval gates in Operational AI Readiness: different consequences deserve different levels of sign-off, and “the gate should fit the consequence, not the other way around” (Bakos, 2026, Ch. 9). Chapter 8 of the same book makes a related point about AI systems themselves: changes to models, prompts and tools need a record, an owner and a way back if something breaks.
ENISA’s own recommendations point the same way. Ireland’s National Cyber Security Centre, quoted in the paper, puts the first steps simply: “Now is the time to review asset inventories, prioritise patching discipline, and assess exposure to unsupported components” (ENISA, 2026, p. 8). For defenders, ENISA recommends treating “every environment as potentially already compromised”, aiming to detect and respond to intrusions within single-digit minutes, and moving faster on segmentation and zero-trust designs (p. 15). It also reminds readers that security basics still decide outcomes, citing research that in 75 percent of breaches, logs existed that should have flagged the attack but were not acted on (p. 11). By that count, in most breaches the warning was already sitting in the logs.
At the European level, ENISA suggests using existing law, including the NIS2 directive, the Cyber Resilience Act and the AI Act, to make sure risks from the most capable models are assessed and reduced, and it floats EU-wide benchmarks for testing the cyber capabilities of advanced models (p. 14).
What it means for you
For most people, the practical message is unglamorous. Turn on automatic updates for your phone, computer, browser and apps, and install updates promptly when asked. Pay attention to devices that no longer receive updates at all, such as old routers, phones and smart home gadgets: ENISA singles out end-of-life products as a growing problem (ENISA, 2026, p. 5). If a device cannot be patched, its window of exposure never closes.
- For everyone: are automatic updates switched on for your devices? Which of your devices no longer gets security updates, and can it be replaced?
- If you rely on a service provider: ask how quickly it applies critical security fixes, and whether that has changed this year.
- For organisations: which fixes can you apply without waiting for a meeting, and who decided that? If the answer is “none” or “nobody”, that is your Authority Gap.
- For leaders running AI systems: do changes to your models, prompts and connected tools have an owner, a record and a tested way to roll back?
What we still don’t know
This is ENISA’s assessment, not incident data. The paper does not measure harm caused by AI-discovered vulnerabilities in the EU, and it does not show that any specific attack used these techniques. The “500 per day” figure is one unnamed organisation’s account, relayed by ENISA. The 15-minute and 72-minute figures come from industry sources ENISA cites, which we have not independently checked. ENISA itself calls its recommendations initial and expects to refine them with member states. And the paper is clear that AI helps defenders too; it focuses on the risks by design.
So the question I’ll be watching: if the defenders get the same AI as the attackers, does the window close again, or does the advantage go to whoever is allowed to act first?
Go further. Operational AI Readiness is FairByDesign’s free field manual for AI that is already live. Chapter 9, Approval Gates for High-Impact Actions, shows how to match sign-off to consequence so urgent changes are not stuck behind the wrong gate, and Chapter 8, Change Management for Models, Prompts, Tools, and Retrieval, covers keeping control when things change quickly. Read it free, or test your own setup with the free readiness assessment, focusing on domains 5 and 6.
Next step: ENISA says it will refine these recommendations in line with the European Commission’s Action Plan on Cybersecurity and AI, which the Commission presented the same day as ENISA’s paper, and we will report what changes. For analysis of AI security guidance as it is published, join Mercury Intelligence for $10 a month (or $100 a year; see what is included).
FairByDesign is reader-funded. If this helped, get the free Mercury Brief or see ways to support. Support never buys editorial control.
References
Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live. FairByDesign. https://fairbydesign.org/operational-ai-readiness/
European Union Agency for Cybersecurity. (2026, July 7). ENISA’s view on cybersecurity in the frontier AI era. ENISA. https://www.enisa.europa.eu/publications/enisas-view-on-cybersecurity-in-the-frontier-ai-era (PDF: https://www.enisa.europa.eu/sites/default/files/2026-07/ENISA%20view%20on%20cybersecurity%20in%20the%20frontier%20AI%20era_en_0.pdf)




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