Summary
- Forescout researchers identified 99 vulnerabilities across ten embedded TCP/IP stacks, including 66 previously unknown findings.
- Of the new vulnerabilities, 52% were rated high or critical severity; 54 were reported to vendors and maintainers, with 24 fixed at publication.
- The research used approximately ten hours of automated analysis and 80 hours of human validation, excluding the work needed to develop the specialised research system.
Cybersecurity researchers have identified 66 previously unknown vulnerabilities in networking software used by industrial and connected devices, employing commercially available artificial intelligence models to accelerate technical investigations that previously required months of specialist analysis.
Forescout identified 99 vulnerabilities across ten embedded TCP/IP stacks during approximately 90 hours of research. Of those findings, 66 were previously unknown, while the remaining 33 involved vulnerabilities that researchers had already identified during earlier investigations.
The company’s Vedere Labs team combined specialist AI software, conventional static analysis tools, and human validation to examine components used in networked equipment. The research, published on 12 November, investigated whether generally available AI models could identify security weaknesses in software relevant to critical infrastructure at substantially lower cost.
Forescout rated 52% of the newly discovered vulnerabilities as high or critical severity. It reported 54 of the 66 new findings to vendors and maintainers, while 24 had been fixed at the time of publication.
Embedded networking software creates shared exposure
TCP/IP stacks provide the software functions that enable devices to communicate over networks. They handle protocols used to send, receive, and interpret information, making them important components in equipment ranging from industrial controllers to connected sensors and building management systems.
Manufacturers frequently incorporate networking software developed by third parties rather than implementing every communications function themselves. A vulnerability within a shared component can consequently affect products supplied by several organisations, even when the finished devices perform different industrial tasks.
Forescout examined ten stacks, including Contiki, FNET, PicoTCP, Nucleus NET, wolfIP, lwIP, and Ethernut, alongside three proprietary implementations.
Seven of the targets had featured in its earlier Project Memoria research, while three had not previously been examined as part of that work. The selection allowed the team to compare its AI-assisted findings with known vulnerabilities as well as investigate additional software.
Across the ten stacks, the system identified 33 vulnerabilities already known from earlier research and 66 new findings. The previously undiscovered issues included weaknesses affecting core network protocol implementations, as well as functions associated with particular applications or services.
The researchers estimated that vulnerabilities within the examined components could affect hundreds of products from numerous manufacturers. That assessment concerns potential exposure through shared software dependencies, not a verified count of vulnerable devices currently operating in industrial environments.
Actual exposure depends on the version and configuration of the networking software, the services enabled on a device, its network accessibility, and any additional protections implemented by the manufacturer.
Establishing which finished products contain affected components therefore requires investigation beyond the discovery of vulnerabilities in the original code.
Automated analysis reduces the research workload
Forescout developed a specialised AI research system capable of coordinating conventional analysis tools with language models and further technical checks. The approach uses automation to identify suspicious code patterns and help researchers determine which findings require deeper investigation.
The system incorporates tools including Semgrep and Joern, with AI models assisting in examining possible vulnerabilities and preparing them for validation.
Approximately ten hours were spent on automated analysis across the targets, followed by roughly 80 hours of human-led verification. The researchers compared this 90-hour process with approximately nine months of manual code analysis undertaken during earlier work.
That comparison suggests a substantial improvement in research productivity, although the investigations were not identical controlled experiments. The software targets, tools, and analysis arrangements differed, meaning the figures should not be treated as a universal benchmark for vulnerability research.
The comparison also excludes the work required to develop the specialised AI system. Forescout had to prepare the analysis tools, build testing infrastructure, establish suitable rules, and refine the instructions used by the models before carrying out the reported investigation.
The research team spent approximately $935 on automated static analysis, or about $2.58 per candidate finding across the 362 candidates processed. Including estimated model expenditure for additional stages, the overall model spend was approximately $1,200 to $2,300.
Those figures represent the marginal cost of using the completed system, not the full cost of the research. They do not include the salaries of the specialists involved, the development of the automation framework, or all supporting infrastructure.
Forescout also found that analysing proprietary binary software was more expensive and time-consuming than examining available source code. Decompiled material created additional uncertainty, requiring more effort to distinguish genuine vulnerabilities from misleading analysis results.
AI still misses important classes of weakness
The research demonstrated limitations in the automated approach as well as improvements in speed. Of the previously identified vulnerabilities within the re-examined software, the system rediscovered approximately 40%.
That result indicates that the AI-assisted process was not a complete replacement for established vulnerability research methods. A substantial proportion of known issues remained undetected, even though the system discovered additional weaknesses missed during earlier investigations.
Memory safety issues accounted for approximately 49% of the new findings, including problems involving out-of-bounds memory access and buffer handling.
Integer-related issues represented 27%, while input validation weaknesses accounted for 16%. Control flow and logic problems formed the remaining 8%, reflecting the system’s greater effectiveness at identifying code patterns associated with memory corruption and protocol parsing.
Implementation-specific weaknesses, such as unpredictable protocol state or flawed random number generation, can be harder to identify through generic static analysis. They may require a detailed understanding of how a particular network protocol behaves under different operating conditions.
Forescout’s findings therefore demonstrate that automation can accelerate some forms of vulnerability discovery without establishing that AI can comprehensively assess the security of embedded networking software.
Remediation depends on manufacturers and operators
Discovering vulnerabilities is only the beginning of the remediation process. Software maintainers must confirm the findings, develop corrections, and distribute the relevant updates to manufacturers using their components.
The equipment manufacturers may then need to integrate those fixes into firmware, test them against the finished product, and make validated updates available to customers.
Of the 66 newly identified vulnerabilities, Forescout reported 54 to vendors or maintainers because it assessed them as relevant to current software versions. The remaining findings were not reported because the affected code was considered absent from the latest versions.
At publication, 24 reported vulnerabilities had been fixed, while additional findings remained under investigation or awaiting remediation.
This process can be particularly complicated for industrial equipment that remains operational for many years. Devices may depend on software whose original maintainer no longer provides updates, while the manufacturer responsible for the finished product may have ended technical support.
Forescout identified seven vulnerabilities associated with an unsupported wireless access point, including hardcoded administrator credentials. Such weaknesses can remain present where the equipment cannot receive a software fix.
Industrial operators also face constraints on maintenance. Equipment supporting production, building services, healthcare, or critical infrastructure may require carefully managed shutdowns before firmware can be updated.
Even where a patch is available, organisations must establish whether applying it could affect the operation of connected systems or introduce compatibility problems.
Asset visibility becomes part of vulnerability response
Forescout recommends assessing vulnerabilities according to their accessibility, exploitability, and operational consequences, rather than treating every finding as presenting the same immediate risk.
Where updates are available, operators can prioritise remediation based on the systems affected and the consequences of compromise. When equipment cannot be patched, restricting unnecessary services and isolating network connections may reduce exposure.
Such measures require accurate information about the assets operating within an environment. An organisation may know the manufacturer and model of its industrial equipment without knowing which TCP/IP stack, firmware components, or third-party libraries are incorporated into each device.
That lack of software visibility can make it difficult to establish whether a newly disclosed vulnerability affects a particular facility. A security advisory may name a software component that is present within several products but never appears in the operator’s ordinary equipment inventory.
Manufacturers can support customers by identifying affected product versions and providing clear guidance about patches and temporary mitigations. Where devices are no longer supported, operators must decide whether the remaining exposure can be managed or whether replacement is necessary.
Forescout has not established that attackers are exploiting all the newly identified vulnerabilities. Its research demonstrates weaknesses in foundational networking software and the ability of AI-assisted tools to accelerate their discovery.
The findings may increase the volume of vulnerability reports facing manufacturers, while industrial operators continue to depend on the availability of validated patches and accurate information about their equipment.
The immediate work involves completing coordinated disclosure, distributing available fixes, and identifying devices that contain affected components. Longer term, faster discovery will place additional pressure on the processes through which software security updates reach equipment operating in industrial and critical infrastructure environments.












