Using generally available AI models, Forescout Research – Vedere Labs identified 66 previously unknown vulnerabilities in ten TCP/IP stacks. These network components are used in operational technology (OT), the Internet of Things (IoT) and critical infrastructure.
Forescout-Studie: KI verkürzt Schwachstellenanalyse von neun Monaten auf 90 StundenForescout Research – Vedere Labs has examined whether generally available AI models can uncover vulnerabilities in open-source software and proprietary binaries on a limited budget. The team developed an AI-assisted vulnerability research platform and applied it to ten widely used TCP/IP stacks. Seven had featured in earlier Vedere Labs research, while three were analysed for the first time.
Researchers validated 99 vulnerabilities, including 66 previously unknown flaws. Forescout rated 52 per cent of the new findings as high or critical in severity and reported 54 to vendors and maintainers. Coordinated disclosure is ongoing for some findings. According to Forescout, the analysis took 90 hours, compared with around nine months for a comparable manual investigation. Automated static analysis cost a total of US$935.
Daniel dos Santos, Forescout’s Vice President of Research, said AI does not remove the need for experienced vulnerability researchers, but substantially reduces the work involved. This could enable both defenders and attackers to examine complex embedded software more quickly and cheaply. For critical infrastructure operators, the findings increase the pressure to identify, prioritise and contain risks promptly, particularly where older devices are difficult to update.
Key findings
The AI-assisted analysis took 90 hours in total. Around ten hours were spent on independent AI analysis, followed by approximately 80 hours of dynamic validation. A comparable manual investigation had previously taken about nine months. Across the ten TCP/IP stacks, researchers validated 99 vulnerabilities: 66 were previously unknown, while the other 33 were known flaws rediscovered by the system.
Fifty-two per cent of the 66 new vulnerabilities were rated high or critical. The AI-assisted system also rediscovered 40 per cent of previously known vulnerabilities. Forescout says this demonstrates that the method can uncover flaws even in software that has already undergone manual research. Researchers reported 54 new findings to vendors and maintainers; coordinated disclosure is still under way for some cases.
Forescout puts the total cost of static analysis across the ten stacks at US$935 and the average cost at US$2.58 per finding. Analysis of proprietary binaries took twice as long and cost around six times more than source-code analysis. However, the results indicate that AI-assisted research can also be applied to firmware whose source code is not publicly available.
Previous Vedere Labs research suggests that vulnerabilities in foundational TCP/IP components could affect hundreds of products from multiple vendors. Where the same component is widely deployed, a single flaw can create risks across the technology supply chain.
Infobox: Study at a glance
- 10 TCP/IP stacks were examined; seven had featured in earlier research.
- 99 vulnerabilities were validated, including 66 new findings.
- 52 per cent of the new findings were rated high or critical.
- The AI-assisted analysis took 90 hours, compared with around nine months for a comparable manual investigation.
- 54 new vulnerabilities were reported to vendors and maintainers.
- Static analysis cost US$935, with an average of US$2.58 per finding, according to Forescout.
Implications for critical infrastructure
TCP/IP stacks enable connected devices to communicate over networks. They are embedded in products used in OT, IoT, medical devices, network equipment and other cyber-physical systems. Because the same software may be used in products from several manufacturers, a vulnerability in one component can expose systems across the wider supply chain.
Remediation is particularly difficult when vulnerable software is embedded in devices that remain in service for decades. Even when a patch is available, it must pass from the software developer through firmware and product manufacturers before reaching the organisation operating the device. Discontinued or unsupported products may never receive an update.
Forescout CEO Barry Mainz said that faster vulnerability discovery does not change the operating constraints faced by asset owners. Legacy systems, operational requirements and devices that cannot easily be patched or taken offline remain common. Forescout therefore argues that risk containment should be treated as a priority alongside patching. Organisations need to know which devices are connected, determine which vulnerabilities pose a real operational risk in their environment and apply controls to contain that risk.
Recommendations for risk mitigation
As AI accelerates vulnerability discovery for researchers and attackers alike, Forescout recommends that organisations prioritise risks according to reachability, exploitability and potential operational impact. Assessments should take account of each device’s function, network exposure, existing safeguards and the possible consequences for operations, rather than treating every vulnerability as equally urgent.
Organisations should also reduce the unnecessary exposure of connected devices. Where possible, services and interfaces that are not required should be disabled or restricted, including FTP, Telnet, IPv6 and web-based administration interfaces. Devices that cannot be patched should be isolated in tightly controlled network segments, with only the communications required for operation permitted.
When fixes are available, organisations should review and accelerate patching processes for embedded devices to keep pace with faster vulnerability disclosure. Security should also be built into procurement: Forescout recommends requiring suppliers to define support lifetimes, publish vulnerability disclosure policies, provide software bills of materials (SBOMs) and demonstrate ongoing security testing.
Forescout further recommends that software vendors and maintainers integrate AI-assisted security analysis into development and release processes and prepare to handle a growing volume of AI-generated vulnerability reports. As vulnerability discovery becomes more scalable, operators need comprehensive visibility into where vulnerable software is deployed and enough context to assess its operational importance and select appropriate risk-reduction measures. The full report is titled “Finding Critical Infrastructure Bugs with Off-the-Shelf AI.”
Infobox: Recommendations for operators
- Prioritise risks by reachability, exploitability and operational impact.
- Disable or restrict unnecessary services and administration interfaces.
- Isolate unpatchable and unsupported devices in controlled network segments.
- Accelerate patching for embedded devices when fixes are available.
- Set procurement requirements for support lifetimes, disclosure policies, SBOMs and ongoing security testing.
Frequently asked questions
What did the Forescout researchers examine?
Forescout Research – Vedere Labs tested whether generally available AI models could find vulnerabilities in open-source systems and binaries relevant to critical infrastructure, using a limited budget. The team developed an AI-assisted platform and applied it to ten widely used TCP/IP stacks. Seven had been included in earlier research, while three were analysed for the first time.
What are TCP/IP stacks?
TCP/IP stacks are foundational software components that enable connected devices to communicate over networks. They are used in OT, IoT, medical, network and other connected devices.
How many vulnerabilities were found?
Researchers validated 99 vulnerabilities. Of these, 33 were already known and rediscovered by the AI-assisted system, while 66 were new. The team reported 54 new vulnerabilities to vendors and maintainers. Forescout rated 52 per cent of the new findings as high or critical.
How much faster was the analysis with AI?
The investigation took 90 hours. Around ten hours were spent on independent AI analysis and approximately 80 hours on dynamic validation. A comparable manual analysis had previously taken around nine months.
Why do the findings matter for critical infrastructure?
TCP/IP stacks are used in many OT, IoT, medical, network and other connected products. Vulnerabilities in these foundational components can affect downstream products across the software supply chain. Older and unsupported devices can be difficult or impossible to patch.
How can operators reduce their exposure?
Forescout recommends maintaining an inventory of connected assets, prioritising vulnerabilities according to their actual risk and selecting appropriate remediation measures. Where devices cannot be patched or taken offline promptly, compensating controls such as network segmentation and isolation can help contain risk.
Source: Forescout Technologies, Inc., press release of 12 November 2026; Forescout Research – Vedere Labs, Finding Critical Infrastructure Bugs with Off-the-Shelf AI. ca.finance.yahoo.com

