Security researchers at Forescout have demonstrated that artificial intelligence can significantly lower the barrier to entry for discovering vulnerabilities in embedded software. By leveraging AI models alongside traditional security scanning tools, the team identified 99 vulnerabilities across 10 different TCP/IP stacks, which are essential components for network communication in industrial, medical, and networking equipment. Of these findings, 66 were previously unknown, with over half classified as high or critical in severity.
The research team developed a specialized system that utilizes Claude models to automate the identification and validation of security flaws. The process begins with code scanners that flag suspicious patterns, followed by AI agents that analyze whether the code is reachable and exploitable. In the final stage, the system attempts to execute the software to trigger the vulnerability. To guide this process, the researchers created 32 scanning rules based on past investigations, focusing on common issues such as buffer overflows and improper network data handling.
The efficiency of this automated approach represents a stark contrast to manual analysis. While the researchers previously spent roughly nine months conducting manual code reviews during a 2020-2021 project, the AI-driven system completed its automated inspection in approximately 10 hours, followed by 80 hours of validation. The total cost for the automated inspection was under $1,000, with total model costs for the entire process ranging between $1,200 and $2,300.
The system proved particularly adept at identifying memory errors within code responsible for processing network traffic. However, it faced greater challenges when detecting complex software logic or design flaws, such as the use of predictable values where randomness was required. Additionally, analyzing compiled software without access to source code proved more difficult and expensive, costing six times more and taking twice as long as analyzing open-source code due to the need to reconstruct the program's structure.
The findings have direct implications for critical infrastructure, where legacy devices often remain in operation long after they have reached their end-of-life. In one instance, the researchers analyzed a discontinued wireless access point from a company that is no longer in business. The system uncovered seven device-specific vulnerabilities, including hardcoded administrator credentials and a buffer overflow flaw in the firmware upload process. Because the manufacturer no longer exists, there is no entity available to provide patches or security updates for these devices.
Siemens has confirmed that certain building automation controllers within its APOGEE, Desigo, and TALON product lines are affected by the discovered vulnerabilities. Other vendors are currently investigating their respective products to determine the extent of the risk. The research team has reported 54 of the newly discovered vulnerabilities to the relevant developers and maintainers, and some projects have already released patches to address the issues.
The names of three networking stacks remain withheld while disclosure processes are ongoing. Because many of the affected devices are difficult to patch or cannot be easily taken offline, the researchers emphasize that organizations must shift their focus toward risk containment rather than relying solely on patching. This research highlights how AI is accelerating the pace of vulnerability discovery, forcing security teams to prioritize mitigation strategies for legacy systems that may never receive a formal security update.





