IoT & Edge IoT & Edge Security

Forescout Finds 99 Flaws in Embedded TCP/IP Stacks With AI-Assisted Analysis

Forescout’s Vedere Labs says it identified 99 vulnerabilities across ten embedded TCP/IP stacks in about 90 hours using commercially available AI models as part of a broader security-analysis workflow. The findings include 66 previously undisclosed flaws, underscoring the potential to examine foundational networking software more quickly—not the ability of AI alone to replace security researchers.

Forescout reported that 52% of the vulnerabilities were rated high or critical severity and that 29% affected core networking components. Embedded TCP/IP stacks provide communication functions used by products ranging from industrial controllers and building automation systems to network equipment and consumer electronics. A flaw in a shared component can therefore warrant investigation across multiple products, although its presence in a stack does not by itself establish that every product using it is exploitable.

A combined analysis workflow

The work builds on Forescout’s earlier Project Memoria research into embedded TCP/IP implementations. The company says that effort took about nine months, compared with roughly 90 hours for the latest AI-assisted analysis. It estimates spending $1,200 to $2,300 on AI models, around 100 times less than the estimated cost of equivalent manual research.

Those cost figures have an important limitation: they exclude the initial development of the research framework, specialized analysis rules and supporting infrastructure. They describe the operating expense of using an established system, not the total investment required to create one.

Researchers combined static-analysis tools, AI reasoning, reverse engineering and dynamic validation, with human review of results. The analysis environment used Ghidra, Semgrep and Joern alongside Claude AI models. This approach was applied to open-source implementations and proprietary software binaries, which can be the only practical target when device manufacturers integrate third-party software without access to its source code.

Forescout says binary-based static analysis was about six times more expensive and took more than twice as long as analysis of source code. The system also had limitations in finding known issues: it rediscovered 33 of 82 previously documented vulnerabilities in the research set, or about 40%. It performed better on memory-safety and protocol-parsing weaknesses than on implementation-specific logic or design flaws.

Implications for connected equipment

The affected implementations include Contiki, lwIP, PicoTCP, FNET, wolfIP and Siemens Nucleus NET. Forescout says it reported 54 of the 66 previously undisclosed vulnerabilities to software vendors and maintainers; disclosure processes for some proprietary stacks remain ongoing.

The researchers identified affected Siemens building automation controllers in the APOGEE, Desigo and TALON product families. They also reported vulnerabilities in a discontinued wireless access point whose manufacturer is no longer operating. The examples point to different remediation challenges: supported products may receive fixes through firmware updates, while devices from defunct suppliers can be difficult or impossible to patch.

Operators still need to establish whether a particular device is affected. Exposure can depend on the software version, configuration, enabled features and network accessibility. The research does not show that all devices using the identified stacks are vulnerable or that the flaws are being exploited.

For industrial and building systems, patching can require compatibility checks, operational testing and a maintenance window. Devices may also remain deployed for many years. Manufacturers and system integrators can reduce uncertainty by keeping software component inventories, monitoring upstream advisories and maintaining processes for vulnerability disclosure and firmware support. Operators can identify which assets contain affected components, assess whether vulnerable functions are reachable and consider compensating controls when an update cannot be applied promptly.

Faster discovery raises response demands

Forescout recommends prioritizing vulnerabilities according to exploitability, network exposure and operational impact, limiting unnecessary services and isolating equipment that cannot be updated. These measures are especially relevant where immediate remediation would disrupt essential operations, but they do not substitute for determining which products and versions are affected.

The research’s broader significance is the potential change in the economics of examining embedded software, including binaries that are harder to assess than source code. AI-assisted techniques may let researchers investigate components that have received less scrutiny. The same accessibility could also make vulnerability discovery available to a wider range of actors, although the study does not establish that malicious actors have used this specific system.

For connected-device makers and infrastructure operators, the practical challenge is to keep vulnerability response commensurate with faster discovery. Knowing what software is deployed, tracking supplier disclosures and planning firmware maintenance are central to managing risk in products built on shared networking components.