An autonomous AI agent called Cipher, built by Causal Security, has found four previously unknown vulnerabilities in EPICS — the Experimental Physics and Industrial Control System, a decades–old open–source toolkit used to build and remotely control distributed SCADA systems. It took 3 hours and 49 minutes. The most serious flaw, CVE–2026–114595, is rated CVSS 9.8 and could allow remote code execution; three further high–severity bugs could cause memory corruption, process crashes or information disclosure. The US Cybersecurity and Infrastructure Security Agency has issued an advisory, and organisations using EPICS should update to EPICS Base 7.0.11, which contains the fixes.
EPICS is specialised software, used in around 90 major scientific, university and industrial projects, from particle accelerators to the ITER fusion reactor. Most UK businesses will never run it. The reason this story matters more widely is what it says about how quickly weaknesses in operational technology can now be found. According to Causal Security co–founder Shubham Raj, Cipher’s initial investigation “required no human intervention”, followed by a single instruction authorising deeper testing in a secure environment. Raj also said exploiting the bugs would not require deep OT expertise. For any UK business with industrial, manufacturing or building control systems, the comfortable assumption that such systems are too obscure or specialised for most attackers to understand is weakening. The defence that still works is the one CISA’s advisory points to indirectly: an attacker first needs network reach to the vulnerable services, and well–segmented control networks deny it.
What was found, and what CISA advises
EPICS was originally developed by Los Alamos and Argonne National Laboratories. It sits between operators and physical equipment, handling measurements, commands and control functions in large technical facilities. That position is exactly why vulnerabilities in it matter: a flaw in software that controls physical equipment has consequences beyond data.
Cipher found four vulnerabilities. The critical one, CVE–2026–114595, is an integer overflow bug rated CVSS 9.8 that could allow a remote attacker to achieve code execution. The other three — CVE–2026–114596, CVE–2026–114597 and CVE–2026–114598 — are high–severity bugs that could cause memory corruption, process crashes or information disclosure.
CISA’s advisory warns that successful exploitation “could allow an attacker to cause a denial–of–service condition, corrupt memory, disclose sensitive information, or potentially execute arbitrary code”. Its recommended action is straightforward: update to EPICS Base version 7.0.11, which includes the fixes. For organisations running EPICS, that is the immediate priority.
The circumstances of the discovery are what make it notable. According to Shubham Raj, Cipher’s “initial investigation required no human intervention, followed by one instruction authorizing deeper exploitability testing in a secure environment”. He described Cipher as “a fully autonomous super hacker that can reason across the technology stack” — a company’s description of its own product, and worth reading as such. What is not in dispute is the outcome: four new vulnerabilities in a mature, widely studied piece of industrial software, found in under four hours and serious enough to prompt a CISA advisory.
This is not Cipher’s first such finding. Earlier in 2026 it found multiple critical attack paths in power–grid systems used across North America and Europe, prompting a separate CISA advisory on 8 October 2026. The pattern is of AI–driven discovery reaching further into critical infrastructure software, and finding things that years of human attention had not.
Operational technology has long benefited from a kind of accidental protection: the software is obscure, the protocols unusual, and the knowledge needed to understand it rare. Raj’s comment cuts against that directly: “Someone with conventional exploitation skills could investigate them, and increasingly capable AI systems could help develop exploits without requiring specialized OT expertise.” If AI tools can find serious flaws in specialised industrial software in hours, the gap between “a flaw exists” and “someone can use it” is narrowing for OT just as it has for ordinary IT. That does not mean every control system is about to be attacked. It means obscurity should no longer be counted as a defence, and that the controls which do not depend on it — separation of control networks, no direct internet exposure, careful management of remote access, and prompt updates where vendors provide them — matter more than they did.
From national laboratories to a four–hour AI review
The sequence below follows the reporting. Some steps are undated in the coverage and are marked accordingly.
The detail worth noticing is the secure environment. The deeper testing was carried out under controlled conditions, which is how responsible vulnerability research is meant to work: findings are confirmed safely, reported, and fixed before details become widely known. The same tools, however, are not limited to responsible researchers. As AI systems become more capable, both defenders and attackers can use them to examine software that once demanded rare expertise.
What limits exposure for control systems
The chart below is an indicative planning model — not measured data — ranking common measures by how much they reduce the risk from vulnerabilities in operational technology and control software, including flaws that have not yet been discovered.
The top two bars reflect the one prerequisite the reporting highlights: to exploit these EPICS flaws, an attacker needs to be able to connect to the vulnerable services on a facility’s control network. That becomes possible if networks are not properly segmented or another connected system provides a path in. Separation and the absence of direct internet exposure therefore protect against these flaws and against flaws nobody has found yet — which, given how quickly AI is now finding them, is a significant advantage.
The bottom bar is the assumption this story undermines. Obscurity was never a strong defence, but for operational technology it often felt like one: few attackers understood the systems well enough to cause harm. When AI can find serious flaws in specialised software in hours, and exploitation needs less specialist knowledge, that comfort largely disappears.
One critical flaw among four
Of the four vulnerabilities Cipher identified, one was rated critical. CVE–2026–114595 carries a CVSS score of 9.8 out of 10. The other three, CVE–2026–114596, CVE–2026–114597 and CVE–2026–114598, complete the set addressed in the CISA advisory. The reporting does not give a separate score for each of those three, so we do not guess one here.
A CVSS score of 9.8 sits close to the top of the scale. In general terms, scores at that level are reserved for flaws that can be exploited over a network with little complexity and serious consequences for confidentiality, integrity and availability. For a control system, the practical meaning of “availability” can be physical: equipment that stops, misbehaves or has to be shut down safely.
That is why the network–reach condition matters so much. A critical score describes how bad things could be if an attacker can reach the vulnerable service. Whether an attacker can reach it is something each operator controls through network design. A well–segmented control network does not make the flaw disappear — the update is still needed — but it shrinks the set of people who could ever try to use it.
How exposed is a typical operator? A readiness score card
The score card below rates common positions on operational technology and specialised software. It is a judgement framework for discussion rather than an audit result, and it applies to any organisation running control systems, building management, specialist manufacturing kit or laboratory equipment — not only EPICS users.
Most organisations will recognise themselves somewhere in the middle. The “Partial” rows are where effort usually pays off fastest: patching that happens but without a schedule, remote access that works but is never reviewed, and monitoring that exists on the office network but stops at the door of the plant room or laboratory.
What it costs to close the gap
The figures below are indicative ranges for a smaller UK organisation with some operational technology on site — a manufacturer, a laboratory, a facility with building management systems or a multi–site business with connected equipment. They are planning estimates, not quotations, and they will vary with the number of sites, devices and suppliers involved.
| Measure | Indicative one–off cost | Indicative ongoing cost | What it addresses |
|---|---|---|---|
| Asset and connection inventory of OT and specialist systems | £1,500–£5,000 | Reviewed annually | Knowing which software and versions you run, and what talks to what |
| Network segmentation between office IT and control systems | £3,000–£15,000 | £50–£300 per month in management | Removing the network reach that attacks like this depend on |
| Secure remote access for suppliers | £1,000–£6,000 | £30–£200 per month | Replacing always–on connections with logged, approved sessions |
| Patch testing and update scheduling | £500–£3,000 to set up | Staff or partner time each cycle | Applying fixes such as EPICS Base 7.0.11 without disrupting operations |
| Network monitoring covering control segments | £2,000–£10,000 | £100–£600 per month | Spotting unexpected connections early |
| Penetration test of the boundary between IT and OT | £3,000–£12,000 | Repeated annually or after change | Confirming that segmentation works in practice |
| Cyber Essentials certification for the office estate | A few hundred pounds upward, depending on organisation size | Annual renewal | Baseline controls on the systems most likely to provide a way in |
Set against those figures is the cost of an incident affecting physical operations: downtime, damaged equipment, safety investigations, contractual penalties and the time spent restoring systems that were never designed to be rebuilt quickly. For many organisations, a single day of halted production would exceed the cost of the segmentation work.
Two ways to approach the same risk
Reactive: patch when an advisory lands
Common, but increasingly fragile
Resilient: assume flaws exist and limit reach
Built for faster, AI–assisted discovery
The difference between the two columns is not the number of flaws in the software. Both organisations run the same products. The difference is how many people could ever reach those flaws, and how quickly the organisation can respond once a fix is published.
How ready are smaller organisations?
The gauge below is our indicative estimate of how prepared a typical smaller UK organisation with connected equipment is for AI–accelerated vulnerability discovery in its operational technology. It is a judgement based on the patterns in the score card above, not a survey figure.
A score of 31 reflects a common picture: office IT is reasonably managed, often with Cyber Essentials or similar controls, but the equipment side — building management, production lines, lab instruments, specialist controllers — was installed by suppliers, connected for convenience and rarely revisited. The good news is that the steps which move the needle most are well understood and do not require specialist OT security tools to begin.
Before buying anything, map which computers, accounts and suppliers can connect to your control systems, building management and specialist equipment. If the answer includes ordinary office PCs, the guest Wi–Fi or the open internet, segmentation is your first priority. If you run EPICS, update to EPICS Base 7.0.11 as CISA advises. For everyone else, the same principle applies: a flaw an attacker cannot reach is a flaw that buys you time.
Analysis: what changes when discovery takes hours
The headline number is the 3 hours and 49 minutes. Specialised control software has traditionally been protected, in part, by the effort required to understand it. A researcher looking for weaknesses in EPICS would need to learn how the toolkit is structured, how its components communicate and how it behaves in the environments where it runs. That learning curve deterred casual attackers and slowed even skilled ones.
Cipher’s result suggests the learning curve is flattening. Its initial investigation needed no human intervention, and a single instruction was enough to authorise deeper testing in a secure environment. Shubham Raj of Causal Security has framed the implication directly: AI makes it far faster to find serious flaws in operational technology, and it reduces the specialist knowledge required to exploit them. The concern is not that one research tool found four bugs; it is that the same class of capability is becoming generally available.
Three consequences follow for anyone responsible for connected equipment.
The gap between disclosure and attack narrows
When finding a flaw takes weeks, defenders often have time between a public advisory and widespread attempts to use it. When AI can analyse a published fix and the software it changes in a short time, that window shrinks. The practical response is to shorten your own patch cycle where you can, and to make sure that systems you cannot patch quickly are not reachable by people who might try.
“Nobody would bother” stops being a plan
Many organisations have quietly relied on the idea that their specialist systems are too niche to interest attackers. That idea always had weaknesses — attackers scanning for anything reachable do not care what it is — but AI removes much of the remaining effort barrier. Our earlier coverage of why security through obscurity no longer holds in the AI era makes the same point from a different angle.
Network design does most of the work
The reporting is explicit that exploiting these EPICS flaws requires network access to the vulnerable services, which can happen where networks are poorly segmented or where another connected system offers a path in. That makes segmentation the single most valuable control, because it protects against the four known flaws and against the ones still to be found. It is also a control that a smaller organisation can implement with ordinary networking equipment and good advice, without specialist OT products.
There is a positive reading too. The same speed that worries defenders also helps them. Cipher’s findings were tested in a secure environment, reported, and addressed in EPICS Base 7.0.11, with a CISA advisory to tell operators what to do. Earlier in the year, its power–grid findings led to a CISA advisory on 8 October 2026. AI–assisted research, done responsibly, means flaws get found and fixed by defenders before attackers stumble on them. The organisations that benefit are those able to act on advisories quickly.
Does this affect UK small businesses?
Very few UK SMEs run EPICS. It is used in large scientific, university and industrial projects, not in a typical office or shop. The story matters for smaller organisations for a different reason: it shows how quickly AI is now able to examine specialised software of every kind.
Many smaller businesses run equipment that falls into the same broad category — building management systems, access control panels, CCTV recorders, production machinery with remote support, environmental controls in laboratories or cold stores, and industrial routers fitted by suppliers. Each runs software that few people understand in depth. Each is often connected to the same network as the office PCs, sometimes with a supplier connection that nobody has reviewed since installation. Our report on ransomware targeting building automation systems shows how that kind of equipment is already in attackers’ sights.
The practical steps are the same whatever the equipment: know what you have, keep it updated where the supplier allows, separate it from the general network and control who can connect to it. Cyber Essentials gives a solid baseline for the office systems that most often provide the way in, and a targeted penetration test can confirm whether the boundary between office and equipment is as solid as it looks on paper.
At a glance
| Item | Detail |
|---|---|
| What was found | Four previously unknown vulnerabilities in EPICS |
| Found by | Cipher, an AI system from Causal Security |
| Time taken | 3 hours and 49 minutes; initial investigation needed no human intervention |
| Human input | One instruction authorising deeper testing in a secure environment |
| Most severe flaw | CVE–2026–114595, CVSS 9.8 |
| Other flaws | CVE–2026–114596, CVE–2026–114597, CVE–2026–114598 |
| Fix | Update to EPICS Base 7.0.11, as advised by CISA |
| Who uses EPICS | Around 90 major scientific, university and industrial projects, including particle accelerators and ITER |
| Key prerequisite | Network access to the vulnerable services on the control network |
| Earlier finding | Critical attack paths in power–grid systems; CISA advisory issued 8 October 2026 |
| Main lesson | Segment control networks and patch promptly; obscurity is no longer a defence |
For related coverage, see our reports on AI finding ways around incomplete security patches, the maximum–severity GitLab remote code execution flaw, the critical PHP remote code execution vulnerability, the end of security through obscurity and ransomware aimed at building automation systems.
Build a baseline that keeps attackers off your network
Most attacks on specialised equipment start with a foothold in ordinary office systems. Cyber Essentials certification covers firewalls, secure configuration, access control, malware protection and patching — the controls that stop that first step. We help UK organisations prepare, fix the gaps and pass.
Talk to us about Cyber Essentials CertificationFrequently asked questions
Not sure what can reach your equipment?
We review UK organisations’ networks, identify where office systems and specialised equipment overlap, and put Cyber Essentials controls in place so the obvious ways in are closed before AI–assisted attackers find them.
Talk to us about Cyber Essentials CertificationAI did not create the flaws in EPICS; it found them faster than anyone had before. The fixes are published, the advisory is clear, and the most effective defence — keeping attackers off the networks where control systems live — is within reach of organisations of every size.
Cloudswitched
London-based managed IT services provider offering support, cloud solutions and cybersecurity for SMEs.
