Forescout’s Vedere Labs pointed an off-the-shelf AI harness at ten TCP/IP stacks and found 99 vulnerabilities in 90 hours.
The same team had analysed seven of those stacks by hand in an earlier project. That work took roughly nine months of code analysis. The AI run took 90 hours and somewhere between $1,200 and $2,300 in model spend — the researchers put the static-analysis portion at $935, an average of $2.58 per finding. They estimate the whole exercise cost about 100 times less than a senior vulnerability researcher doing the same work manually.
Sixty-six of the 99 were new. Fifty-two percent were rated critical or high severity. Twenty-nine percent sat in the core of the stacks rather than in application-level protocol handling, which is the part reachable before anything authenticates.
The speed is the headline. The consequence is the story: the cost of finding exploitable flaws in the embedded software inside industrial equipment just collapsed, and the cost of fixing that equipment did not move at all.
The Economics of Finding Bugs Changed. The Economics of Patching Did Not.
An asymmetry sits underneath every OT security programme. Finding a vulnerability and remediating one have always both been expensive, and roughly in proportion. That proportion just broke.
A researcher — or an attacker — can now run an automated analysis across a class of embedded software for the price of a few hours of consulting. The operator on the other side still needs a vendor to issue firmware, a maintenance window to apply it, and in regulated environments a safety recertification before the equipment goes back into service. Nine months of research became 90 hours. Eighteen months of patch lead time became eighteen months.
Forescout’s own conclusion on the worst case is worth quoting directly: AI-enhanced research can now “quickly and inexpensively uncover dangerous vulnerabilities for which no vendor will be able to provide a patch.”
Why a TCP/IP Stack Is the Worst Place to Have a Flaw
The stacks tested are not obscure research code. Contiki, Ethernut, FNET, PicoTCP, lwIP, wolfIP and Siemens’ Nucleus NET sit inside low-power devices, Ethernet-attached industrial equipment and real-time operating systems in safety-relevant gear. Forescout’s earlier work on the same family of flaws suggests they reach hundreds of products from dozens of vendors.
A TCP/IP stack is the code that turns packets arriving on a wire into data the device acts on. It runs before the login screen. It runs before any access control the vendor layered on top. A memory-corruption flaw there is reachable by anything that can send the device a packet, which in a flat plant network means anything on that network.
Of the 99 findings, 48 involved memory corruption, 27 involved integer handling, 16 were input-validation problems and eight were control-flow or logic issues. That distribution is the signature of code written for constrained hardware, often a decade or more ago, by people solving a different problem.
One case from the research deserves its own paragraph. One stack was analysed as it shipped inside the firmware of a discontinued wireless access point, running as part of a third-party operating system. The access point is discontinued. The company that made it no longer exists. The harness still found seven device-specific vulnerabilities in it, including a buffer overflow via firmware upload and hardcoded administrator credentials that cannot be changed for either the web or the shell interface.
There is no patch coming for that device. There is no vendor to issue one. And equipment in exactly that condition is sitting in plants, substations and building risers across the country right now.
The Four Reasons OT Patch Management Stalls
Four reasons, and only one of them is about effort.
The vendor is gone or the product is end-of-life. No firmware is coming because nobody is writing firmware. The access point above is the pure case, but every plant has a handful.
Safety recertification. Equipment governing a process with a safety case cannot take a firmware change without revalidation. That is a documented, scheduled, sometimes externally witnessed process. “Apply the patch this week” is not a sentence that survives contact with it.
Change windows measured in quarters or years. A continuous process does not stop because an advisory landed. Treatment trains, furnaces, compressors and lines have planned outages, and those are when firmware moves.
There is no update path at all. Some embedded devices were never designed to be updated in the field. The firmware is the firmware.
OT patch management, done honestly, is therefore not a programme for applying patches quickly. It is a programme for deciding which of the flaws you cannot fix actually matter, and what you put in place around the ones that do. Forescout’s own guidance to asset owners says it in the same order: assess and prioritise by reachability, exploitability and operational impact; reduce exposure by turning off services you do not need; isolate devices that cannot be patched and allow only strictly required traffic; and speed up patching where a fix does exist.
Three of those four are not patching.
Even When a Patch Exists, It Has to Reach the Device
There is a second gap underneath the first, and it is the one that makes this a supply-chain problem rather than a vendor-diligence problem.
These stacks are components. A maintainer fixing a flaw in Contiki or lwIP publishes a fix for the component. That fix then has to be picked up by whichever real-time operating system embeds it, integrated by the equipment vendor who embeds that OS, built into a firmware release, published, and finally installed by the operator. Forescout states the condition plainly: these supply-chain flaws likely affect hundreds of products from dozens of vendors, and stack vendors’ patches must still reach affected devices through the supply chain.
Every link in that chain is a separate company with its own release schedule and its own commercial reasons to deprioritise a device it sold eight years ago. One of the stacks in the study was Siemens’ Nucleus NET, which carried 20 vulnerabilities from the earlier manual research — the same vendor whose industrial product families appear in CISA ICS advisories month after month. A well-resourced vendor that actively issues advisories is the good case, and even there the fix has to traverse the chain.
So the operator’s position is this: the number of known flaws in their equipment is about to rise sharply, because finding them got cheap. The share of those flaws that reaches them as installable firmware will not rise at anything like the same rate. The gap between the two is the risk that has to be managed by other means.
What the Standards Say When Patching Is Impossible
This is not a gap the governing standards ignore. IEC 62443 is built around zones and conduits precisely because it assumes some equipment in a zone cannot be hardened directly. The control is the boundary around the device and the monitoring across it, not the device itself.
CISA’s guidance across its ICS advisories follows the same shape: minimise network exposure so control systems are not reachable from the internet, put them behind firewalls and isolate them from the business network, and use secure methods such as a VPN for remote access while recognising that the VPN is only as secure as the devices connected to it.
Nothing in either framework assumes a patch will arrive. Both assume some of it will never be fixed and ask what you have built around it. A compensating control is not a lesser option in these documents. It is the expected answer.
What Detection Can and Cannot Do Here
Being precise about this matters more on this topic than on most, because a post about unpatched equipment invites the suggestion that a monitoring product makes the problem go away. It does not.
IntelliThreat does not patch embedded firmware. There is no mechanism by which a detection platform fixes a buffer overflow in a TCP/IP stack. The flaw stays.
IntelliThreat does not scan OT devices for these CVEs. Active scanning of fragile control equipment is how people knock plants over, and an asset-inventory platform built for that job is a different product.
What it does is watch the network around the equipment. Self-Healing MNDR and Autonomous SIEM inspect every packet and log, 24/7, with full metadata and payload analysis across endpoints, servers, ICS/OT and IoT. Exploiting a stack-level flaw means sending the device traffic. That traffic is observable even though the flaw is not fixable.
Concretely, that means a controller reached from a host that has never spoken to it before, a management or firmware-upload session from an unexpected source, malformed or unusual protocol traffic aimed at a device that normally sees a narrow and repetitive pattern, and the scanning and enumeration that precedes all of it.
Containment in 2.9 seconds. Detected at T+0, correlated at T+1.4s, contained at T+2.9s. On equipment with no fix available, time-to-containment is the number that matters, because prevention is off the table by definition.
Plain-language alerts. The operator who has to act on this is frequently a controls engineer, not a security analyst. The verdict arrives in English.
98% less alert noise in the first 30 days. A plant team with no dedicated security staff cannot work a queue.
Private AI. The AI runs inside your own environment and your operational data never leaves your tenant. On a post about devices you are not permitted to send anywhere, this is not a footnote: the monitoring must come to the equipment, not the other way around.
Where it still stops. Serial links that never touch Ethernet are invisible to a network sensor. Sensor placement governs coverage — a sensor on the business network does not see an isolated process segment. An attacker operating entirely inside a controller, generating no network traffic, is outside what any network-based approach can observe. And none of this is a substitute for the segmentation work.
Thirty Days: Inventory, Segment, Watch
In rough order of risk removed per hour spent:
- Inventory by software, not just by asset. The question is no longer “what devices do we have” but “what embedded stacks and third-party components are inside them.” Ask vendors for a software bill of materials. The answers you cannot get are themselves findings.
- Flag everything end-of-life or orphaned. Any device whose vendor is gone, or whose product line is discontinued, belongs on a list of its own. That list will never be remediated by patching.
- Turn off services nobody uses. Forescout names FTP, Telnet, IPv6 and web administration interfaces specifically. Unused services on embedded devices are free attack surface.
- Segment the unpatchable into their own restricted zone, allowing only the traffic they genuinely need — the zone-and-conduit model, applied to the devices that need it most.
- Put monitoring on the conduits. Segmentation without observation tells you a boundary exists but never tells you when something crossed it.
- Decide what a normal day looks like for each device. A controller that talks to three hosts and one historian has a describable pattern. Deviation from it is the signal. If nobody can describe the pattern, that is the first thing to fix.
- Plan replacement for the orphans. It is the only real remediation. Budget it on the assumption that more flaws in those devices will surface cheaply from now on.
Steps one and two are the ones most teams have never completed, and they are the ones that determine whether any of the rest is possible.
Common Questions
What is OT vulnerability management? It is the process of identifying, assessing and mitigating vulnerabilities in operational technology — the control systems, industrial equipment and embedded devices that run physical processes. It differs from IT vulnerability management in a way that changes the whole programme: in IT, remediation usually means applying a patch. In OT, a large share of findings have no patch available, no maintenance window within the quarter, or no update mechanism at all, so the work shifts from remediation to prioritisation and compensating controls.
What is ICS security? ICS security protects industrial control systems — PLCs, HMIs, RTUs, historians and specialised controllers. Availability and safety outrank confidentiality, patch windows are rare, and equipment lifecycles run decades. Compromise is measured in physical consequence rather than records exposed.
What does OT cybersecurity mean in practice for a plant with no security staff? Knowing what is connected, keeping the unpatchable equipment in a restricted zone, watching the traffic in and out of that zone, and having something that acts in seconds when the traffic changes — without requiring an analyst on site to interpret it.
What is ICS security monitoring? Watching traffic to and from control equipment for configuration changes, unexpected management sessions and command activity that should not be happening. It is the control that keeps working when patching is not available, which on this equipment is most of the time.
If AI is finding these flaws, can AI not also fix them? Not on the device. A fix has to be written by whoever maintains the stack, integrated by the equipment vendor, shipped as firmware, and installed in a maintenance window. Every one of those steps is a human and commercial process, and for discontinued products with defunct vendors the chain is broken at step one. The research cut the cost of finding flaws, not the cost of remedying them.
See It on Your Own Network
Start a 30-day trial and watch IntelliThreat triage your real traffic. 30 days free. No credit card required. Plans for teams of 10 to 10,000.
Or request a demo and we will walk your control network before you connect anything.
Related reading: critical infrastructure cybersecurity and industrial and OT cybersecurity.