Forescout: AI models found 99 flaws in 10 TCP/IP stacks for about $2,300
Forescout says a Claude-based system confirmed 99 vulnerabilities in networking code used in industrial, medical and network devices, at a model cost of roughly $1,200 to $2,300.
By Mr Elite Published 7 min read
- Severity
- High
- Status
- Reported by outlets
- Sources
- 1
- Last checked
The short version
- Forescout confirmed 99 vulnerabilities across 10 TCP/IP stacks; 66 were previously unknown, and 52% of those were rated high or critical.
- Help Net Security reported model costs of about $1,200 to $2,300, with roughly 10 hours of automated analysis and 80 hours of validation.
- Siemens has identified affected APOGEE, Desigo and TALON building automation controllers; other vendors are still investigating.
- The system found 33 of 82 known vulnerabilities used for comparison, and was weaker on logic and design flaws.
- Some affected products come from vendors that no longer exist, so no patch can be expected.
What happened
Forescout researchers have confirmed 99 vulnerabilities across 10 TCP/IP stacks using a system built on Claude models and existing security tools, Help Net Security reported on 16 November 2026. A TCP/IP stack is the software component that lets a device communicate over a network.
The team wanted to test whether generally available AI models could find flaws on a limited budget. They examined open-source code and compiled software whose source code was not available, an approach the researchers noted attackers could also use.
Of the 99 confirmed vulnerabilities, 66 were previously unknown, and 52% of those new ones were rated high or critical. The affected software is used in industrial equipment, medical devices, network equipment and other connected products.
Siemens has identified affected building automation controllers in its APOGEE, Desigo and TALON product lines, according to the report. Other vendors continue to investigate.
The team reported 54 of the new vulnerabilities to developers and maintainers because it believed they still affected the latest versions. Several projects have issued fixes. Investigations remain open for three stacks whose names have been withheld during disclosure.
Why it happened
This is a research disclosure rather than a single incident, so there is no one root cause. The flaws were in the networking code itself. The researchers' scanning rules targeted problems such as copying more data than a memory buffer can hold and mishandling information received over a network.
The pipeline worked in stages. Code scanners flagged patterns that might indicate a vulnerability. AI agents then judged whether an attacker could reach the suspect code and exploit it. Finally, the system tried to run the software and build a test that triggered the issue, with researchers able to step in and guide it.
Thirty-two scanning rules came from records of earlier investigations, and the team reviewed and refined them first. The process produced 362 initial findings, reduced to 312 candidates after duplicates were removed. Testing and review confirmed 99.
How this kind of failure works
Networking code is a rich hunting ground because it takes input from strangers. Every packet arriving at a device is data the code did not write and cannot trust. If the code mishandles that data, an attacker may be able to crash the device or run their own instructions on it.
A buffer overflow is the classic example. Picture a form with a name box that holds 20 characters. If the software accepts a 200-character name and writes it all down, the extra text spills into neighbouring boxes and overwrites whatever was there. In a device, those neighbouring boxes may control what the program does next.
Pattern-matching tools are good at spotting this kind of mistake, which is why the system did best on memory errors in code that processes network traffic. Flaws of design are harder. Forescout said predictable values meant to be random were among the problems the system found difficult to detect, because nothing in the code looks obviously broken.
Compiled software adds a further hurdle. Without source code, tools must rebuild a usable picture of the program from machine instructions, which can introduce inaccuracies and false alarms. Forescout said this made inspection about six times as costly and roughly twice as slow.
Who is affected
The affected software is used in industrial equipment, medical devices and network equipment. Siemens has named APOGEE, Desigo and TALON building automation controllers. Forescout has not published a full product list, and other vendors are still investigating, so the complete set of affected devices is not yet known.
Operators can start by asking their vendors whether their products use any of the affected stacks. Three stack names have been withheld, which makes it harder to check independently until disclosure finishes.
The hardest cases are products with no vendor left. Forescout described firmware from a discontinued wireless access point made by a company that had ceased operating. The system found seven device-specific flaws there, including administrator credentials that could not be changed and a buffer overflow in the firmware upload process.
How to defend against it
Tick each step as you check it. Your progress is saved on this device.
What to learn from this
The concept: Compensating controls for unpatchable systems
Patching is the normal answer to a vulnerability, but it needs a vendor who still ships fixes. When that vendor has gone, or the device cannot be taken offline, you must reduce risk by other means. Segmentation, access limits and monitoring are those means.
Forescout's chief executive, Barry Mainz, said containing risk rather than patching must be the priority for such systems. The same idea applies to old industrial controllers, unsupported operating systems and medical equipment with long service lives.
The research also shows why the problem may grow. If finding flaws costs a few thousand dollars, old firmware is cheaper to examine, and more of it will be examined.
When a device cannot be patched, limit who and what can talk to it.
Quick check
Sign in to earn 10 XP for a right answer
A building controller has a known flaw, and its maker has closed. Which response best fits the lesson?
With no patch possible, containment limits what an attacker can reach, and monitoring helps spot attempts. Waiting for a patch that will never come leaves the risk unchanged.
Timeline
- 2020 to 2021 Project Memoria, an earlier manual investigation, studied TCP/IP stacks; Forescout says it ran 18 months overall and covered 14 stacks.
- Before 16 November 2026 Forescout built and ran its AI-assisted system against 10 stacks, producing 362 initial findings and 312 candidates.
- Before 16 November 2026 Testing and review confirmed 99 vulnerabilities, 66 of them previously unknown.
- Before 16 November 2026 The team reported 54 new vulnerabilities to developers and maintainers; several projects issued fixes.
- 16 November 2026 Help Net Security published the findings; Siemens had identified affected APOGEE, Desigo and TALON controllers.
What is not known yet
- Which three TCP/IP stacks remain under open investigation, as their names have been withheld.
- Which other vendors' products are affected, since investigations are continuing.
- Whether any of the newly found flaws have been exploited in the wild; the report does not say.
- Whether fixes will ever be issued for products whose vendors no longer operate.
Questions people ask
What did Forescout find?
Forescout said it confirmed 99 vulnerabilities across 10 TCP/IP stacks, 66 of them previously unknown. Of those new ones, 52% were rated high or critical. The software is used in industrial, medical and network devices.
How much did the AI analysis cost?
Help Net Security reported that automated code inspection cost $935, and total model costs ranged from about $1,200 to $2,300. Automated analysis took about 10 hours, followed by roughly 80 hours of validation. These figures cover applying the finished system, not building it.
Which products are affected?
Siemens has identified affected building automation controllers in its APOGEE, Desigo and TALON lines. Other vendors are still investigating. Forescout has not released a full list, and three stack names are withheld during disclosure.
Can these flaws be patched?
Several projects have issued fixes for reported flaws. Some affected products, however, come from vendors that no longer operate, so no patch can be expected. For those devices, Forescout's chief executive said containing risk should be the priority.
Sources
- Help Net Security: AI uncovers flaws in discontinued devices with no vendor left to patch them