AI finds flaws in devices with no manufacturer left to repair them

AI finds flaws in devices with no manufacturer left to repair them

A discontinued wireless access point had seven device-specific vulnerabilities, including administrator credentials that could not be changed and a buffer overflow in the firmware update process. The company that made it no longer exists: there is no vendor to prepare a patch. Forescout found the flaws while testing an AI-assisted vulnerability analysis system.

The research matters for two distinct reasons. AI reduces the time and cost needed to search network software for bugs; at the same time, it can uncover flaws in deployed devices that are no longer maintained. For organizations that depend on those devices, discovering the problem does not mean they can fix it.

From nine months to about 90 hours

Forescout applied its system to ten TCP/IP stacks, components that enable devices to communicate over networks. Seven had already been analyzed in previous work, and three were new to the team. The process identified 99 vulnerabilities: 33 that were already known and 66 that were new. The comparison with the earlier manual analysis refers to about nine months of code analysis; in the new work, automated analysis took about ten hours and validation about 80.

The cost of the models for static analysis was $935. Forescout estimates that the total model cost, including dynamic testing, came to between $1,200 and $2,300. The comparison does not include designing the system, preparing environments to run the software, or developing and tuning the rules and instructions. Those figures therefore describe the cost of using an already-built tool, not the full cost of creating it.

MeasureProject Memoria, 2020-20212026 study
ApproachManual analysis with traditional toolsAI system supervised by researchers
Stacks analyzed14 across several phases10 in one phase
Results in the stacks reviewed71 vulnerabilities99: 33 rediscovered and 66 new
Code analysis timeApproximately nine monthsAbout 10 hours of automated analysis and 80 hours of validation

The difference in speed does not mean the system found and confirmed every flaw without intervention. The process generated 362 initial findings; after duplicates were removed, 312 candidates remained, which had to be reviewed and tested. Only 99 were confirmed as vulnerabilities. Human and dynamic validation are part of the result, not a dispensable detail.

A search tool combining rules and models

The method did not involve asking a chatbot to examine code on its own. Forescout built a specialized system for embedded software, combining static analysis tools with AI agents based on Claude models. First, rules prepared by the researchers flag suspicious patterns; then, the agents assess whether an attacker could reach the code and whether the flaw appears exploitable.

The final phase attempts to run or emulate the software and produce a test that triggers the problem. Researchers can intervene and guide the analysis at this stage. Forescout created 32 analysis rules based on patterns observed in previous research, such as errors when processing data received over a network or copying information beyond a buffer's capacity.

This design also explains some of the limitations. The system was better at locating memory errors in code that processes network traffic; logic or design flaws were more difficult. Cybersecurity Dive reports that Forescout attributed this difference to the use of static rules based on generic patterns: a logic error usually depends more on how each program is implemented.

Analyzing compiled programs without available source code was another challenge. The tools had to reconstruct a representation of the software, which could introduce inaccuracies and false alarms. According to Forescout, this work cost about six times more than analyzing source code and took approximately twice as long.

“Organizations still have legacy systems, operational requirements, and devices that can’t easily be patched or simply taken offline. Containing risk, rather than patching, must be the priority.”

— Barry Mainz, Forescout CEO

The patch does not always reach the device

The potential scope extends beyond the ten stacks examined. Forescout notes that flaws in shared components could affect hundreds of products from dozens of manufacturers, although each case depends on which software the product incorporates and how fixes are distributed through the supply chain.

Forescout reported 54 of the 66 new vulnerabilities to developers and maintainers, considering that they still affected current versions. Its detailed report says that 20 affected open-source stacks and 34 affected proprietary stacks; it also notes that five vulnerabilities had been fixed at the time the report was published. Cybersecurity Dive, meanwhile, reports that 24 of the 54 reported vulnerabilities had already been fixed, 29 were still under investigation, and one affected the Ethernut project, whose developer did not respond. These are counts published with different breakdowns and should not be treated as if they describe exactly the same remediation status.

The access point case illustrates the most severe limitation: the tool can discover flaws, but it cannot bring the manufacturer back. When the device is still needed and cannot be updated or retired, the response is to reduce its exposure. Forescout recommends assessing the flaw's reachability, exploitability, and operational impact; it also suggests restricting unnecessary services and isolating devices that cannot be patched.

Siemens publishes measures for its controllers

The problem is not limited to products without a manufacturer. Siemens identified vulnerabilities in building automation controllers in its APOGEE, Desigo, and TALON lines. Its security advisory lists products that may be vulnerable to denial of service and loss of availability. Updates are available for some models; for others, Siemens says it does not plan to release a fix.

Affected products according to SiemensRecommended measure
APOGEE PXC Compact and PXC Modular, depending on protocol and versionUpdate to V3.5.7 or later, or V2.8.21 or later, depending on the model
TALON TC Compact and TC Modular, BACnetUpdate to V3.5.7 or later
APOGEE MBC and MEC; several Desigo controllers from V2.3 onwardSiemens says it does not plan to patch the listed products

For several vulnerabilities in APOGEE, Desigo, and TALON products, Siemens recommends disabling FTP/Web services; for Desigo controllers affected by another vulnerability, it advises segmenting the network and placing them behind properly configured gateways or firewalls. The advisory also offers updates for specific models. The measure depends on the product: there is no single instruction that applies to the entire list.

More findings, more response work

The change for manufacturers and maintainers is operational: they may receive more vulnerability reports and have less time to assess them as the cost of finding flaws falls. Forescout recommends using AI to discover bugs before attackers do and preparing to handle those reports. The system also demonstrates that automation does not eliminate the need to verify results, especially when analysis starts with binaries or third-party code.

For those maintaining industrial networks or connected devices, the practical takeaway is to review which devices remain exposed, which services they need, and which update or mitigation applies to each model. Forescout warns that a fix to the network stack must reach devices through their manufacturers; a patch available for the component does not guarantee that the installed product will receive it.

The uncomfortable question remains for the team operating the discontinued device: if there is no vendor to fix the firmware, can they isolate it without interrupting a necessary function? In the case of the analyzed access point, there is no manufacturer update to close the seven vulnerabilities found.

Sources: Help Net Security, Forescout, Cybersecurity Dive, Siemens ProductCERT