← Back to all advisories
ELTON exploitability notification
High

Embedded TCP/IP Stacks: 66 New Vulnerabilities in lwIP, Nucleus NET and Eight More

Advisory ID
ELTON-EN-2026-004
Category
Memory safety flaws in packet parsing
Vendor
lwIP, Nucleus NET (Siemens), Contiki, PicoTCP, FNET, wolfIP, Ethernut, and three stacks not yet named
CVE
CVE-2026-120596, CVE-2026-89208, CVE-2026-102073 to 102076. 19 more await CVE IDs and carry Forescout numbers from FSCT-2026-0006 to 0025
Versions analyzed
lwIP 2.2.1, Nucleus NET 5.1, Contiki 3.0, PicoTCP 1.7.0, FNET 4.6.3, wolfIP 1.0, Nut/OS 5.1. Reported because current versions were still affected
CVSS
Up to 8.2 (v3.1, Siemens SSA-333035)
First Published
November 20, 2026
Status
Fixes published for six stacks. No fix for Ethernut. Three stacks under disclosure. No exploitation known to ELTON
MedDevice AI Security MonthlyThe live notifications, worked through on real devices every month. One hour, free, with the recording and slides sent to everyone who registers.Save your seat →
How it is exploited

The chain on an example product

DemoVent V2 is a fictional ventilator, built only to show the path. Its firmware is a real-time operating system with lwIP 2.2.1 from the chip vendor's SDK, it keeps a data port open on the ward network to forward alarms to the central station, and nothing restarts the network task if it stops. The graph is how ELTON reads this notification on a product like it: toggle the entry vector and the preconditions, apply a mitigation, and watch the verdict change. It is the same dependency graph ELTON builds for every product it tests: entry vectors produce conditions, findings require them, and the chain decides which of the four verdicts a finding gets on that device.

PreconditionsNetwork reach to an open TCP port on a device built with lwIP 2.2.1 and out-of-order queuing on, which is the upstream default. No credentials and no user action: the attacker needs only a connection. Forescout says only that the crash works under specific conditions.
PostconditionsThe stack crashes and every network service on the device stops. What follows is decided by the product: a supervised restart brings the network back in seconds, and without one the device stays off the network until someone power cycles it.
VerdictDirectly exploitable. On this example, one connection takes the ventilator off the network. Close the data port to the ward network and it becomes an unexploitable weakness there. Ship firmware with the lwIP fix and it becomes a mitigated weakness; a supervised restart mitigates only the lost alarms.
How ELTON shows the impact

The finding on the digital twin

The chain above is one vulnerability on one example product. The disc below places the same finding on the digital twin of a whole release: the entry vector it comes in through, the paths it opens, and the findings chained behind it.

4,187no pathUSBBLUETOOTHDIGITAL TWINA malformed packet enters at the network stackTESTLINK5GETHERNETTestLink boxDirectly exploitableConditionally exploitable, chained behind it
Where it entersThe TCP/IP stack handles every packet before any other code on the device. On the twin that is every network interface, shown here as Ethernet.
What it reachesA crash or a hang in the stack takes every network service down with it. Where a flaw writes past a buffer, the path can continue into the firmware.
What the fix changesUpdate the stack where a fix exists. Where none is coming, close the port or segment the device, and the path is cut before the parser.

ELTON uses this view to demonstrate impact. The finding is placed on the digital twin of your product, the path is drawn from the interface an attacker can reach, and the same picture shows what a fix removes. TestLink then runs the attack along that path on the physical device, so the impact is shown on the product and recorded as evidence. How ELTON works.

Risk Summary

On November 12, 2026, Forescout Vedere Labs published 66 new vulnerabilities in ten embedded TCP/IP stacks: lwIP, Nucleus NET, Contiki, PicoTCP, FNET, wolfIP, Ethernut and three proprietary stacks it has not named yet. An AI harness built on generally available models found them, along with 33 that were already known, in about 90 hours. Forescout rates 52% of the new findings critical or high, and 17 of them sit in the core of the stack: IP, TCP, UDP, ICMP and Ethernet handling.

A TCP/IP stack handles every packet a device receives, before any login, certificate check or application logic. Most of the published flaws need no credentials and take one packet, a few fragments or one connection. Where Forescout states the result, it is usually a crash, a hang or a denial of service, and in four cases the whole stack stops. The rest are out-of-bounds reads and overflows, and the overflows that write are the class that became remote code execution in Forescout's earlier Project Memoria.

Medical devices sit inside the affected population, and Forescout names them among the products that embed these stacks. A manufacturer rarely chooses its stack. It arrives inside a real-time operating system, a chip vendor's SDK or a wireless module: lwIP ships in SDKs such as Espressif's ESP-IDF and STMicroelectronics' STM32Cube, and Forescout describes Nucleus as used in safety-critical devices such as anesthesia machines and patient monitors.

ELTON is not aware of any reported exploitation. Fixes are published for six of the ten stacks, Ethernut has none, and 29 findings in the three unnamed stacks are still in disclosure. A fix in a stack reaches a fielded device only when its manufacturer ships firmware. ELTON rates this High, and Critical on any product where a network outage or a restart interrupts therapy or alarms.

Technical Details

Forescout repeated its Project Memoria research with AI. Project Memoria ran for 18 months in 2020 and 2021, largely by hand, and produced almost 100 CVEs in 14 stacks. The 2026 harness pairs Claude models with Semgrep, Joern and Ghidra in five stages: it prepares the target, scans it with 32 static analysis rules, checks that the suspect code is reachable, checks that it is exploitable, and then verifies the finding on a built or emulated target. Static analysis of all ten stacks took about ten hours and $935, or $2.58 per static finding. Dynamic validation, supervised by the researchers, took about 80 hours more.

The funnel ran from 362 static findings to 312 unique candidates to 99 confirmed vulnerabilities: 48 memory corruption, 27 integer issues, 16 input validation and eight control flow or logic. The harness rediscovered 33 of the 82 vulnerabilities already known in these stacks and added 66. Forescout reported 54 of the 66, the ones it believes still affect current versions.

Stack and version analyzedNew findings reportedStatus on November 12, 2026
lwIP 2.2.14Fixed in the lwIP source tree
Nucleus NET 5.1 (Siemens)5Fixed by Siemens, advisory SSA-333035. No fix planned for several legacy controllers
Contiki 3.04Fixed in Contiki-NG
PicoTCP 1.7.05Fixed in PicoTCP-NG
FNET 4.6.33Fixed
wolfIP 1.03Fixed
Ethernut (Nut/OS) 5.11No response from the maintainer. No fix
Three proprietary stacks, not yet named7, 10 and 12Under disclosure. Details withheld

The published flaws a device builder should read closely:

FlawWhat the attacker sendsResult
lwIP TCP
FSCT-2026-0006
A connection to any open TCP port, under specific conditions. The out-of-sequence merge loop in tcp_receive() copies a FIN flag from a queued segment without checking that the segment still fits the receive windowThe stack crashes
lwIP MQTT client
FSCT-2026-0007
A crafted PUBLISH message from the broker, delivered as a chained bufferThe lwIP main thread loops forever and every network service freezes
lwIP TFTP server
FSCT-2026-0009
Repeated read or write requests that the server rejects. Each one leaks a timerThe timer pool runs out and the process aborts
lwIP SMTP client
FSCT-2026-0008
From, to or subject fields longer than 240 characters, passed in by the applicationHeap overflow past a 256-byte buffer, reaching a function pointer
Ethernut TCP
CVE-2026-120596
One TCP segment to any open port, carrying an unknown option type with a length of zeroThe TCP receive thread never returns and all TCP processing stops. No fix
Nucleus NET web service
CVE-2026-102073
Input whose length the service does not validate, with no authenticationReads device memory or stops the service. CVSS 8.2
Nucleus NET FTP and web
CVE-2026-102074 to 102076
Input the FTP service, the multipart handling or the upload code does not validate, with no authenticationStack overflow and denial of service. CVSS 7.5 each
FNET fragment handling
FSCT-2026-0023
Three crafted IP fragments, handled before checksum verification or socket lookupHeap overflow
PicoTCP reassembly
FSCT-2026-0017
A crafted series of IP fragments from the adjacent networkHeap overflow in reassembly
Contiki RPL
FSCT-2026-0015
One ICMPv6 routing message from a node on the same linkStack overflow of 16 bytes

In the lwIP source, the merge loop is compiled only when out-of-order queuing is on (TCP_QUEUE_OOSEQ), which upstream lwIP enables by default. The product's own lwipopts.h decides, and some vendor SDK examples turn it off. Forescout says only that the crash works under specific conditions. The appendix of its report lists all 25 published flaws, and CVE IDs for the 19 that carry FSCT numbers are pending at MITRE.

The count is a floor. Forescout reports that the harness is strong on memory safety in parsing code and misses logic and design flaws, such as predictable initial sequence numbers, DNS cache poisoning and TCP state machine errors.

The marginal cost is now small. With the harness built, each new target took hours and a few hundred dollars at most, and closed firmware did not stop it: binary targets cost about six times more to analyze than source and still produced 34 of the 54 reported findings. One of them was firmware from a discontinued access point whose maker no longer exists, where the findings include hardcoded administrator credentials and an overflow in the firmware upload.

Which Devices Are Affected

A device is affected when all of these hold:

  • Its firmware, or the firmware of a module inside it, includes one of the ten stacks at a version without the fix.
  • The vulnerable part is compiled in and in use: TCP with out-of-order queuing, IP fragment reassembly, an MQTT, DNS or DHCP client, a TFTP, FTP or web service, IPv6 neighbor discovery or RPL routing.
  • The interface is reachable by an attacker: a hospital VLAN, a Wi-Fi or mesh radio, a service port, or a broker or server the device connects to, where an attacker controls it or can pose as it.

Where it hides in a medical product:

ComponentTypical stackWho can reach it
Device firmware on a microcontrollerlwIP or another small stack, supplied with the RTOS or the chip vendor's SDKAnyone on the device's network segment
Wi-Fi, Ethernet or cellular moduleA stack inside the module's own firmware, often a fork of lwIPThe wireless network or the wired segment
Low-power sensor or wearable on a meshContiki-class stacks with 6LoWPAN and RPLAny radio neighbor on the link
Legacy product on a commercial RTOSNucleus NET, or one of the proprietary stacks still in disclosureThe clinical network

Outside the seven named stacks: products whose networking is the Linux kernel's own, and builds that leave the vulnerable module out. An lwIP build with TCP_QUEUE_OOSEQ set to 0 does not contain the merge loop, and a product with no MQTT client or TFTP server cannot be reached through one. The three unnamed stacks are proprietary, so a product on a commercial RTOS stays an open question until they are published.

Siemens has already named products: building automation controllers in its APOGEE, Desigo and TALON lines, with updates for some and no fix planned for several legacy models. Other vendors are still investigating.

The SBOM often has no line for the stack. It names the operating system or the SDK, and the stack inside is a vendored copy or a fork with its own version string. Build options decide which protocols exist in the image at all. Three of the ten stacks are still unnamed, so no SBOM search can find them yet.

Finding every affected product across a portfolio is a search on those shared attributes. ELTON runs it on the digital twin it builds for every product during onboarding, before any test runs: a blueprint of the product, deeper than a threat model, with its components, interfaces, software materials and the network each interface listens on. One query across the twins lists every product with an embedded IP stack, the SDK or module it came in, and the access ELTON holds to each one. When the three unnamed stacks are published, the same query runs again.

Exploitability

Most of the published flaws are unauthenticated and remote. Ethernut hangs on one TCP segment, and PicoTCP reads out of bounds on one IPv4 packet that carries IP options. FNET overflows its heap on three fragments, before any checksum or socket check, and lwIP can be crashed by a connection to an open port, under conditions Forescout does not detail.

Others need a position first. The lwIP MQTT flaw needs a malicious broker, the Contiki DHCP flaw needs a rogue DHCP server on the gateway's segment, and the wolfIP driver flaws need the same network segment or control of the USB host the device is plugged into.

The results Forescout states are mostly loss of service: the stack, or a service running on it, crashes or loops. For a device that forwards alarms, reports to a central station or takes orders over the network, that is a safety question, and the design decides the answer. A supervised restart brings the network back in seconds. Without one, the device is off the network until someone power cycles it, and where the stack shares a processor with a clinical function, a crash can reset both.

The overflows that write past a buffer need the closest reading. On a microcontroller with no memory protection, a heap or stack overflow in packet handling is where code execution starts.

Whether a specific product is exploitable depends on which stack and which protocols it was built with, how they are configured, and which ports answer on which network. The answer comes from sending the packet to the device. A component name in an SBOM cannot give it.

How ELTON reaches the verdict: the finding is placed in the product's exploit chain with the entry vector and the conditions it needs. The chain above is the one for the example ventilator, and it decides whether the finding is directly exploitable, conditionally exploitable, an unexploitable weakness or a mitigated weakness on this device. The method is the dependency graph, and it is why a ventilator whose data port is closed to the ward network gets a different verdict from the same flaw than one that answers every connection.

Regulatory View (Section 524B)

Section 524B asks for a plan to monitor, identify and address postmarket vulnerabilities, and for an SBOM that covers commercial, open-source and off-the-shelf components. A TCP/IP stack inside an SDK is an off-the-shelf component, and a reviewer can ask where it is listed.

FDA's premarket cybersecurity guidance lists availability among the security objectives a device's design should address. A flaw that takes a device off the network belongs in the risk assessment wherever a clinical function depends on that connection.

A submission in review that ships one of the named stacks can expect to be asked for its assessment: which version, which protocols are built in, which ports answer, and the test result. Postmarket, the plan needs the same assessment per product, and for a stack with no fix coming, the compensating controls communicated to customers.

The volume is what changes for the plan. A harness that turned nine months of code analysis into 90 hours will be pointed at more components, and a process sized for a few stack disclosures a year has to absorb more of them.

Mitigation

  1. Find the stack in every product, including the copies inside SDKs and wireless modules, with its version and build options.
  2. Update where a fix exists: lwIP from its source tree, and Contiki-NG, PicoTCP-NG, FNET and wolfIP from their repositories. Siemens advisory SSA-333035 covers Siemens' own controllers, so a product that licenses Nucleus NET needs the patch from Siemens. Then ship the firmware.
  3. Turn off what the product does not use: FTP, TFTP, Telnet, web administration, IPv6, and clients such as SMTP that no feature needs.
  4. Take data and service ports off networks that do not need them: closed outside service mode, a device VLAN, firewall rules that allow only required traffic.
  5. Isolate what cannot be patched: products on Ethernut, on a stack past its support date, or on one of the three still in disclosure.
  6. Authenticate the broker and the servers the device connects to with TLS and a pinned certificate, so a crafted reply has to come from the real one.
  7. Supervise the network task with a watchdog, so a crash or a hang ends in a restart, and test that it does.
  8. Record the stack, where it came from and its build options beside the SBOM entry, so the next stack disclosure is a lookup.

Verifying the fix: once the firmware is built, ELTON's agentic pentesting runs the test again against the whole physical product in the lab: the malformed segments and fragments against every open port, the crafted broker reply, and the port check where the patch has not shipped. The test case is the one that found the exposure, so the retest is the same case, and the result is evidence that the stack survives on the shipped build. The reach into the physical device is TestLink, the appliance ELTON puts on your bench or in its own lab. Only ELTON has it, and it carries the pipeline from software into the physical product: the models build the test and TestLink runs it on the real device.

What ELTON Verifies

  1. Inventory from the digital twin and the SBOM: which products carry an embedded TCP/IP stack, which one, at what version, and whether it came in an operating system, an SDK or a module.
  2. Which protocols and services the build includes, and which ports answer on the device's real network interfaces, as shipped and in service mode.
  3. The packets themselves against the real product through TestLink: the published triggers for the stack in use, evidenced, so the disposition rests on a result.
  4. What the device does when the stack stops: whether it restarts, how long the network is gone, and which functions go with it.
  5. Verification that the fixed stack or the interim control is active on the shipped build.
  6. A verdict per product, directly exploitable, conditionally exploitable, unexploitable weakness or mitigated weakness, with the evidence attached for the 524B plan or the submission.

Findings on a covered product arrive in days. The agentic pipeline builds the test once and runs it again after the fix. See exploitability verification and the incident response process.

Disclosure Template

A manufacturer with an affected product sends its customers a disclosure statement: where the exposure sits, the exploit chain, the verdict, the fix and its status, and the test that proves the fix. ELTON builds that statement and the VEX for its customers from platform data. The same statement is available as a blank template to complete by hand, with a finished example, on the disclosure templates page.

Timeline

  • 2021Forescout closes Project Memoria after 18 months: almost 100 CVEs in 14 embedded TCP/IP stacks, found largely by hand.
  • 2025-12General support ends for one of the proprietary stacks now under disclosure. Its extended support ends in September 2027.
  • 2026-10-13Siemens publishes SSA-333035 for five Nucleus NET flaws in APOGEE, Desigo and TALON building controllers.
  • 2026-11-12Forescout Vedere Labs publishes the research: 99 vulnerabilities in ten stacks, 66 of them new, 54 reported.
  • 2026-11-20ELTON exploitability notification published.

References

Find out which products in your portfolio carry one of these stacks.Inventory from the digital twin, the packets sent to the real device, disposition per product.

Questions

Common questions about
TCP/IP stack vulnerabilities.

What did Forescout find in embedded TCP/IP stacks?

On November 12, 2026, Forescout Vedere Labs published 99 vulnerabilities in ten embedded TCP/IP stacks, found by an AI harness in about 90 hours. 66 are new, and 54 of those were reported to vendors and maintainers. The named stacks are lwIP, Nucleus NET, Contiki, PicoTCP, FNET, wolfIP and Ethernut, and three proprietary stacks are still in disclosure. ELTON tracks it as exploitability notification ELTON-EN-2026-004.

Which stacks are fixed, and which are not?

Fixes are published for lwIP, Contiki, PicoTCP, FNET, wolfIP and Nucleus NET, where Siemens advisory SSA-333035 covers its own building controllers. The Ethernut maintainer has not responded, so its TCP options flaw, CVE-2026-120596, has no fix. Three unnamed proprietary stacks with 29 findings are still under disclosure. A stack fix reaches a fielded device only when the device manufacturer ships new firmware.

Why does this matter for a medical device?

The TCP/IP stack handles every packet before any login or encryption, and it usually arrives inside an RTOS, a chip vendor's SDK or a wireless module, so it is easy to miss in an SBOM. Most of the published flaws need no credentials, and the results Forescout states are mostly crashes, hangs and denial of service. On a device that forwards alarms or takes orders over the network, losing the network is a safety question.

What can be done where no fix exists?

Close or segment the ports the product does not need on the clinical network, turn off unused services such as FTP, TFTP, Telnet and web administration, and isolate products built on a stack with no fix. Add a supervised restart so a crashed stack recovers, and where the source is available, correct the parser in the product's own copy. Document the controls for customers.

How does ELTON determine which products are exposed?

The stack is often missing from the SBOM, and the exposure depends on which modules are built in and which ports answer. ELTON reads those from each product's digital twin, then sends the published triggers to the real device through TestLink and records what the device does. Each product gets one of four verdicts with the evidence attached.

Exploitability management for medical devices. FDA §524B methodologyExploitability proven at runtime95% faster than legacy testing Book a Demo →
Platform
OverviewAvoid FDA DeficienciesAvoid Consulting FeesDigital Twin TraceabilityAI MedDevice PentestingFDA Vulnerability VerificationVulnerability ChainingRemediation OptimizationRemote TestLink™Incident ResponseAutomated VEX & MetricsCVSSv4 Migration
Solutions
EnterpriseStartups / SMBs Postmarket SurveillanceIncident ResponseSecurity EngineeringRegulatory AffairsFDA §524BEU MDR/CRAEU REDNIS2IMDRF N60 / N73Japan MHLWAI-Enabled DevicesDevice Modalities
Why ELTON
100% ComplianceFDA TraceabilityFDA Vulnerability VerificationFDA-Qualified CVSSv3CVSSv4 MigrationQMSR Audit Compliance One SolutionTesting ProcessPosture ManagementAI-PoweredELTON vs. Legacy TestingPricing
Resources
FDA Deficiency ListFDA Testing RequirementsFDA Cyber SOPs & TemplatesVulnerability Disclosure TemplatesExample Test ReportRemediation LibraryRegulatory GuidesWebinarsAI NewsletterThe End of Legacy TestingThe AI Vulnerability ExplosionAI Inside the ProductCybersecurity TestingSecurity AdvisoriesWhitepapersIntelligence & Blog
Company
AboutCredentialsLeadershipPartnershipsData SecurityContact Meet ELTON →