Forescout AI system finds 99 flaws across 10 TCP/IP stacks
Forescout researchers used an AI-assisted analysis system to identify 99 vulnerabilities across 10 TCP/IP stacks, including 66 flaws they said were previously unknown. The findings highlight risks in software embedded in industrial, medical and networking equipment, including products that may remain deployed after their manufacturers stop operating.
Of the 66 newly discovered vulnerabilities, 52% were rated high or critical in severity. The research examined both open-source code and compiled software for which source code was unavailable. Siemens has identified affected building automation controllers in its APOGEE, Desigo and TALON product lines; other vendors are still investigating.
The study also illustrates a difficult operational problem for defenders: discovering a flaw does not guarantee that a patch will be available. In one case, researchers analyzed firmware from a discontinued wireless access point whose manufacturer had ceased operating. They found seven device-specific vulnerabilities, including unchangeable administrator credentials and a buffer overflow in the firmware upload process.
How the analysis system worked
Forescout built the system around Claude models and existing security tools, informed by the team’s prior work on embedded software. The process began with code scanners applying 32 rules developed from earlier investigations. The rules looked for patterns associated with defects such as copying more data into a memory buffer than it can hold, or mishandling information received over a network.
AI agents then assessed whether an attacker could reach the suspicious code and whether the suspected weakness might be exploitable. In a final stage, the system attempted to run the software and create a test that triggered the issue. Researchers could intervene and guide the analysis, so the process was not simply an unattended model producing verified vulnerability reports.
That distinction matters when interpreting the results. The system generated 362 initial findings. After duplicates were removed, 312 candidates remained, and testing and human review confirmed 99 vulnerabilities. The researchers also checked the system against 82 known vulnerabilities, rediscovering 33. It performed best on memory errors in code that processes network traffic. Software logic and design issues, including predictable values intended to be random, proved harder to find.
Time and cost depended on validation
Automated analysis took about 10 hours, followed by approximately 80 hours of validation. Forescout compared that work with roughly nine months of manual code analysis during Project Memoria, an earlier investigation conducted in 2020 and 2021. The earlier project lasted 18 months overall and examined 14 networking stacks across several phases. The new study covered 10 stacks, seven of which had been studied before.
The comparison covers applying the completed system to the targets, not the full effort required to build it. Developing the scanning rules, preparing environments in which to run the software and refining instructions for the AI agents required additional work. Automated code inspection cost $935; including estimated spending on tests that ran the software, model costs totaled approximately $1,200 to $2,300.
Analysis of compiled software was more demanding than analysis of source code. Without source, the tools had to reconstruct a usable representation of the program, which can be inaccurate and produce false alarms. Forescout said this work cost about six times as much and took roughly twice as long as source-code inspection.
Legacy devices need compensating controls
Forescout said it reported 54 of the newly discovered vulnerabilities to developers and maintainers because the issues appeared to affect the latest versions. Several projects have issued fixes. Investigations involving three networking stacks remain open, and their names have been withheld during the disclosure process.
For organizations, the research reinforces the need to account for embedded systems that cannot be promptly updated or removed from service. Asset inventories should identify device models, firmware and the business functions they support, while vulnerability response processes should track vendor advisories and remediation status. Where a product is unsupported or cannot be patched, teams may need to reduce exposure through containment and other compensating controls rather than assume a software fix will arrive.
Forescout CEO Barry Mainz said organizations may have legacy systems and operational requirements that make patching or taking devices offline difficult. The practical implication is that vulnerability discovery must be paired with decisions about exposure and operational risk. AI-assisted analysis may increase the speed at which flaws are surfaced, but validation, disclosure coordination and mitigation remain necessary parts of the process.