Security Through Obscurity Is Dead - AI Just Delivered the Fatal Blow

For thirty years, a quiet assumption has propped up a large part of real-world security practice: that a system nobody understands is a system nobody attacks. Not because it was documented policy — almost no organisation has ever written “we are protected because our setup is unusual” into a risk register — but because it was true often enough to be load-bearing. Obscure software, bespoke line-of-business applications, an industrial controller configured by a contractor who retired in 2014, a decade-old open-source library that everyone used and nobody read: these things survived not because they were hardened, but because working out how to break them cost an attacker more time than the target was worth. As of 14 September 2026, senior figures across law enforcement, vendor research and the bug bounty industry are saying publicly that this arithmetic has broken.

The trigger is not a single vulnerability or a single breach. It is a pattern that has become impossible to read any other way. Brett Leatherman, assistant director of the FBI’s Cyber Division, told The Register that the latest AI models were able to find significant vulnerabilities in open-source libraries that had been publicly visible for a decade and were running in roughly 80% of web servers — code that had been through years of community scrutiny and was assumed, reasonably, to be secure. Days after Black Hat 2026, five US federal agencies issued a joint warning that attackers had used AI-generated exploitation scripts to break into internet-exposed Siemens S7 Series programmable logic controllers at water, manufacturing and energy facilities, with the unusually direct line: “this is not a theoretical risk — it is an active threat”. And Dustin Childs, chief bug hunter at Trend Micro’s Zero Day Initiative, speaking the day after Microsoft’s record-breaking September 2026 Patch Tuesday closed 974 CVEs, said that whether security through obscurity is dead “isn’t even an opinion question”.

For a UK SME, the practical translation is narrower and more uncomfortable than the headlines suggest. It is not that AI has invented a new class of attack. It is that the barrier which used to protect your specific, awkward, undocumented systems — the fact that an attacker would have to learn them first — has been substantially removed. If your organisation has never had its systems properly tested, the realistic position today is that an AI-equipped attacker can enumerate your weak points faster than your team can discover them by accident. This article sets out what was said, what the evidence actually supports, where the limits are — and they are real — and what a business with fewer than 250 staff should do about it in the next quarter.

What was actually said, and by whom

Three separate voices made the same argument within days of each other, from three positions that rarely align. That convergence is the reason the story carries weight rather than reading as vendor commentary.

The first was law enforcement. Brett Leatherman of the FBI’s Cyber Division described AI models locating significant flaws in open-source libraries that had been sitting in public repositories for ten years, used in the region of 80% of web servers. The detail that matters is not the percentage but the assumption it demolishes. Open-source code of that reach is the most-reviewed software on earth. “Many eyes make all bugs shallow” was the governing theory, and the practical corollary was that if a decade of researchers, maintainers and downstream integrators had not found a serious flaw, there probably was not one left to find. AI did not need to be cleverer than those researchers. It needed to be tireless, and to read the whole thing at once.

The second was the vendor research community. Dustin Childs runs bug hunting at Trend Micro’s Zero Day Initiative, the largest vendor-agnostic bug bounty programme in the industry, which means he sees the raw inflow of vulnerability reports before anybody else does. His phrasing — that the death of security through obscurity “isn’t even an opinion question” — is notable for what it declines to do. He did not say obscurity has become less effective, or that it is under pressure. He treated it as settled. The timing sharpens it: he was speaking the day after Microsoft shipped fixes for 974 CVEs in a single Patch Tuesday, the largest single release in the programme’s history, which we covered in detail at the time.

The third was the threat intelligence side, and it is the most specific. John Hultquist, one of the more measured voices in the field, pointed out that AI is “excellent at technical troubleshooting, at knowing obscure systems”. His concern was industrial control systems, and his reasoning cuts to the heart of the matter: ICS environments were “largely secured because the expertise was in a handful of people’s heads”. That was never a control. It was a shortage. A shortage of expertise functions as a security boundary right up until the moment the expertise becomes freely available, and a model that has ingested every vendor manual, every forum thread, every conference talk and every decompiled firmware sample is precisely that moment arriving.

How the position hardened through 2026

This did not arrive as a single announcement. It accumulated across the year as a sequence of individually unremarkable events that only make sense read together.

Where obscurity was quietly doing the work

It is worth being honest about how much of the UK SME estate has been relying on unfamiliarity rather than hardening. The categories below are not ranked by how dangerous each system is in isolation. They are ranked by how much of their protection came from the attacker not knowing them — which is exactly the component that AI removes. A system scoring high here is not necessarily insecure; it is a system whose security has never been tested against an adversary who understands it.

Read the bottom bar first, because it explains the top one. Mainstream vendor software — Windows, Microsoft 365, the major browsers, the common firewall platforms — never relied on obscurity in the first place. It is picked over continuously by researchers, criminals and the vendor’s own security teams, which is why 974 CVEs can appear in one month: that number is a measure of scrutiny, not of weakness. Those products are already living in the world the rest of the estate has just been moved into.

The top bars are the ones nobody has ever attacked properly. A bespoke stock control application written for one organisation has, over its lifetime, typically been examined for security by nobody at all. Its protection was never its code quality; it was the fact that an attacker would have to spend a week understanding a system that would net them one mid-sized company. That week is now an afternoon, and a model that has never seen the application can still recognise the familiar shapes inside it: the unparameterised query, the session token that never expires, the file upload that trusts the extension, the admin endpoint protected by nothing but an unguessable URL.

Hultquist’s point about industrial and building control systems deserves particular attention from UK SMEs who assume it does not apply to them. It is not only water treatment plants. A manufacturer with a handful of PLCs on the shop floor, a distribution business with an automated racking system, a landlord or managing agent with a networked building management controller for heating and access control — these are all ICS, and they are overwhelmingly configured once by a specialist contractor and then never touched. They were secured, as Hultquist put it, “because the expertise was in a handful of people’s heads”. That is the exact condition AI dissolves.

The half of the problem AI does not solve

Here is where the story gets more useful than its headline, and where any honest reading has to slow down. If AI has made finding vulnerabilities cheap for attackers, the obvious hope is that it makes fixing them cheap for defenders. The evidence published this week says otherwise, and the numbers are specific enough to plan around.

1Password’s security research team, led by Director of Security Research Keith Hoodlet, took six CVEs disclosed since March 2026 and generated 6,080 candidate patches using OpenAI’s ChatGPT-5.5 and Anthropic’s Opus 4.8. Only 26.0% fully resolved the vulnerability without materially changing application behaviour. Roughly three in four attempts either failed to close the flaw or closed it while breaking something else.

The failure mode is more instructive than the headline rate. Even among the patches that did successfully fix the underlying flaw, roughly 20% also broke other application behaviour — the example given was altering “allow list” logic. That is the worst category of change for a small business to receive, because it passes every test that asks “is the vulnerability gone?” and fails the test nobody runs, which is “does the system still do what the business needs?” An allow list that has silently widened is a new vulnerability wearing the paperwork of a fix. An allow list that has silently narrowed is a Monday morning in which a department cannot work and nobody can say why.

This is the substance behind Katie Moussouris’ intervention, and it is the most practically important comment in the whole story. Moussouris founded the bug bounty consultancy Luta Security and effectively built the modern vulnerability disclosure model; she has spent longer than most watching organisations receive more findings than they can act on. Her assessment has two halves that need to be held together. AI genuinely removes the barrier of needing familiarity with a specific technology stack, letting an attacker enumerate weak spots in systems they have never encountered — that is the new thing, and it is real. But, she stressed, finding bugs was never the hard part. Triage, prioritisation and actually fixing issues remains the bottleneck it has always been.

So the asymmetry is worse than “attackers got faster”. Attackers got faster at the step that was already the easy one, while defenders gained very little on the step that was already the hard one. A 974-CVE Patch Tuesday lands on the same finance director who still has to decide which of those fixes touch systems the business cannot afford to have go down. The 26.0% figure says that handing that decision to a model is not yet a strategy.

Where UK SMEs are most exposed right now

The following is the pattern we see repeatedly across organisations of 10 to 250 staff. The badge indicates how much of the organisation’s protection in that area has historically come from obscurity rather than from a control that would survive contact with an attacker who understands the system.

The four high-rated rows share a characteristic worth naming: in every case, the organisation could not currently answer the question “what would an attacker who fully understood this system be able to do with it?” That question has been theoretical for most SMEs because no such attacker was realistically going to turn up. The events of this year are the argument that the question is now practical, and the only reliable way to answer it is to have someone competent try.

What proper testing costs a UK business

The reasonable objection at this point is budgetary. Penetration testing has a reputation among smaller organisations as an enterprise expense, and that reputation was earned in an era when the alternative — being too small and too obscure to be worth attacking — was a defensible position. The bands below reflect current UK market pricing for a scoped external engagement. They are indicative and vary by complexity, not a quotation.

Two notes on reading that table. First, Cyber Essentials Plus is not a substitute for penetration testing and should not be presented as one. It verifies that five foundational controls — boundary firewalls, secure configuration, access control, malware protection and patch management — are genuinely implemented. It does not ask whether your bespoke application has an authentication bypass. It is the floor, and for many SMEs it is also a contractual requirement for public sector work, which makes it the most straightforward spend on the list.

Second, the patch management control inside Cyber Essentials requires critical and high-severity patches to be applied within 14 days. In a month that produces 974 CVEs from one vendor, that requirement stops being a box to tick and becomes an operational commitment that needs resourcing. The 1Password research is the reason it cannot simply be automated away: a fix applied without regression testing is a change made blind, and roughly one in five successful AI-generated fixes altered application behaviour in the process.

Two postures, and the difference between them

The gap between those columns used to be a question of maturity — nice to close, rarely urgent for a 40-person firm. What has changed is which assumption the left-hand column rests on. It was never “we are secure”. It was “we are not worth the effort of understanding”. Both the FBI’s account of decade-old library flaws and the five-agency PLC warning describe that assumption failing in production.

The story at a glance

This story connects directly to several we have covered in recent weeks, and read together they describe one trend rather than five separate incidents. The mass exploitation of PaperCut NG and MF by an operator running several hundred autonomous AI agents is what “attackers no longer need to understand your stack” looks like at scale. The BlueMoon exploit kit and the Chromium patch gap is the espionage-grade version of the same technique, weaponising public source code faster than the ecosystem could ship the fix. September’s record 974-CVE Patch Tuesday is the volume that makes triage the binding constraint, exactly as Moussouris argues. The rapid exploitation of JFrog Artifactory showed how quickly a fix existing and a fix being installed diverge in practice. And our note on VMO2’s £600m cost pressure is a reminder that the infrastructure and supplier estate you depend on is under its own investment strain at the same moment the attack economics are shifting.

Frequently asked questions

CloudSwitched

London-based managed IT services provider offering support, cloud solutions and cybersecurity for SMEs.