Forescout uses AI to find 66 unknown security flaws in networking software
The Claude-based system made discovery faster and cheaper in Forescout’s study. Human-guided validation and unsupported devices expose the limits of that gain.
Loading page…
The Claude-based system made discovery faster and cheaper in Forescout’s study. Human-guided validation and unsupported devices expose the limits of that gain.
Listen to this story
Forescout’s Claude-assisted workflow confirmed 99 vulnerabilities in 10 TCP/IP software stacks, including 66 flaws not previously known, after researchers used targeted scanning rules to focus the agents on suspicious code. The work suggests AI can accelerate vulnerability discovery, but its reach was selective: it found only 33 of 82 known flaws used for comparison, and most of the 90-hour effort went to validation. Fifty-four new flaws were believed to remain in current versions; fixes are underway for some, while unsupported devices may have no vendor left to patch them.
Researchers rated 52% of the newly discovered vulnerabilities high or critical, with memory errors easier to find than logic and design flaws.
Forescout estimated $935 in automated code-inspection spending, or $1,200–$2,300 including model use for software tests; these figures exclude system-building work.
Analyzing compiled software took roughly twice as long and cost about six times more than working from source code.
Networking software that had already faced expert scrutiny still contained undiscovered security holes. In research disclosed November 12, Forescout said a system combining Claude models and security tools confirmed 99 vulnerabilities across 10 networking software stacks, including 66 previously unknown flaws. The findings also exposed a harder problem: vulnerable equipment whose manufacturer no longer exists.
The targets were TCP/IP stacks: software components that let devices communicate over networks. They are used in industrial equipment, medical devices and network products. Forescout’s test included seven stacks from its earlier research and three it had not previously examined. Researchers rated 52% of the newly discovered vulnerabilities as high or critical severity.
This was not a general-purpose chatbot inspecting code without preparation. Researchers drew on earlier investigations to create 32 scanning rules, then reviewed and refined them. The rules targeted recognizable patterns, such as copying more data than a memory buffer can hold or mishandling information received over a network.
That filtering mattered. The automated process generated 362 initial findings, reduced to 312 candidates after duplicates were removed. Testing and review confirmed 99 vulnerabilities. Separately, the system rediscovered only 33 of 82 known vulnerabilities used for comparison—a reminder that finding new bugs did not mean reliably finding every existing one.
Its strengths followed the rules it was given. Memory errors in code handling network traffic were easier to uncover. Logic and design problems, such as supposedly random values that were predictable, proved harder. Forescout attributed that gap to rules focused on generic code patterns rather than implementation-specific behavior.
Automated analysis took about 10 hours; validation took approximately 80 more. Forescout compared that total with roughly nine months of manual code analysis during Project Memoria, its earlier investigation in 2020 and 2021. Most of the new study’s measured time therefore went into checking potential flaws, not generating leads.
The comparison was not between identical projects. Project Memoria covered 14 stacks across several phases and lasted 18 months overall. The new time estimates cover applying an already-built system to 10 targets. Developing scanning rules, preparing software execution environments and refining instructions for the AI agents required additional work.
Forescout put automated code-inspection spending at $935. Including estimated model spending on tests that ran the software, the total ranged from approximately $1,200 to $2,300. Those are model costs for applying the system, not an all-in price for building and operating the research effort.
Access to source code also made a difference. Inspecting compiled software cost about six times as much and took roughly twice as long. Tools first had to reconstruct a usable representation of the program, a step that could introduce inaccuracies and false alarms.
Researchers reported 54 new vulnerabilities they believed still affected current software versions. Several projects have issued fixes, while investigations remain open for three unnamed stacks. Siemens has identified affected building automation controllers in its APOGEE, Desigo and TALON lines, connecting the software findings to specific equipment families.
One investigation reached firmware from a discontinued wireless access point whose manufacturer had ceased operating. It also uncovered seven device-specific vulnerabilities, including administrator credentials that could not be changed and a buffer overflow in the firmware upload process.
For equipment with no supplier left to provide a patch, discovery cannot lead to a conventional vendor repair. Forescout CEO Barry Mainz argues that containing risk must take priority when devices cannot easily be patched or taken offline.
Loading discussion...
Join the conversation
Explain when containing the risk would be enough.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.