Artificial intelligence helped uncover dozens of previously unknown vulnerabilities in software used in networking equipment, industrial systems, and critical infrastructure. A study by US company Forescout showed that publicly available AI models can reduce software analysis time from roughly nine months to 90 hours.
During the experiment, researchers identified 99 vulnerabilities in ten popular TCP/IP stack implementations—the foundational software that enables network-connected devices to exchange data. Of the issues discovered, 66 were previously unknown.
More than half of the new vulnerabilities—52%—were rated as high or critical severity. A total of 54 identified issues were reported to software developers and maintainers for remediation.
The findings were presented by Forescout Research — Vedere Labs in a report titled “Finding Critical Infrastructure Bugs with Off-the-Shelf AI.” The company announced them on November 12, 2026.
What exactly the AI examined
A TCP/IP stack is a set of software components that enable data transmission between devices over a network. It is used almost everywhere: from computers and routers to medical equipment, industrial controllers, building management systems, and Internet of Things (IoT) devices.
Vulnerabilities in these components are particularly dangerous because the same software code may be used in equipment from different manufacturers. A flaw in foundational networking software can therefore potentially affect many devices and systems at once.
The researchers examined ten popular TCP/IP stacks, including seven previously studied as part of Project Memoria and three new ones. They analyzed both open-source software and closed-source binaries for which source code is unavailable.
The AI was not simply used as a chatbot tasked with finding bugs. The researchers built a dedicated system combining artificial intelligence models, static code analysis tools, and mechanisms for validating the issues discovered.
The automated phase took about 10 hours. Researchers needed approximately another 80 hours for dynamic testing and validation of the findings. In total, 99 potential issues were confirmed as genuine vulnerabilities, 66 of which were previously unknown.
However, the reported 90 hours do not include development of the system itself: building the tools, configuring the testing environment, and preparing instructions for the AI required additional work. The time savings therefore apply to the actual investigation using an already prepared system, rather than the entire technology development cycle.
The automated static analysis cost $935. Including additional testing costs, the researchers estimated total model usage expenses at $1,200–2,300.










