Get DeafLetter
A weekly selection of signals, vulnerabilities and guides. Critical alerts remain optional.
You can unsubscribe at any time. Privacy policy.
On November 13, 2026, Forescout Research Vedere Labs published findings that rewrite the economics of vulnerability research. A system built on Claude models analyzed 10 embedded TCP/IP stacks, uncovering 99 vulnerabilities, 66 of them previously unknown. Fifty-two percent of the new flaws are rated high or critical. Total time: roughly 90 hours, versus nine months for the manual Memoria project analysis (2020-2021). Cost: $935 for the automated inspection, $1,200 to $2,300 including dynamic testing.
The stakes aren't speed alone. It's what happens when AI finds flaws in hardware that will stay in the field for decades, made by defunct vendors or discontinued lines nobody maintains anymore. The discovery accelerates everyone: defenders, but also those hunting entry points.
- Forescout used Claude models with a static scanner, AI agents for reachability and exploitability analysis, and dynamic emulation to generate triggerable proof-of-concepts
- Of the 66 new vulnerabilities, 54 were reported to vendors or maintainers; only 5 had been fixed at publication time
- Forty-nine percent of bugs fall into memory corruption, 27% into integer issues, 16% into input validation
- A discontinued wireless access point firmware from a defunct vendor yielded 7 device-specific vulnerabilities, including unchangeable admin credentials and a buffer overflow in the firmware upload process: no patching possible
How Forescout's System Works
The Forescout system combines three components. A static scanner with 32 rules targeting vulnerable patterns — buffer overflows, integer issues, input validation. AI agents for reachability and exploitability analysis: they determine whether a flaw is network-reachable and triggerable. Finally, dynamic emulation to generate triggerable proof-of-concepts.
The system operates on both source code and compiled binaries. In the latter case, reverse-engineering costs run roughly six times higher. The methodology rediscovered 33 of the 82 known vulnerabilities used for benchmarking, 40% of the total.
The component breakdown shows where flaws concentrate: core IP/TCP/UDP/ICMP/Ethernet at 29%, HTTP/web server at 15%, IPv6/ND/RPL/DHCP at 10% each, Telnet at 7%. The data comes from Industrial Cyber, which confirms Help Net Security's numbers with additional CVSS distribution details.
The Risk Numbers: Severity and Categories
According to Industrial Cyber, the CVSS distribution of the 66 new vulnerabilities is: 47% high (7.0-8.9), 45% medium (4.0-6.9), 5% critical (9.0+), 3% low (below 4.0). Help Net Security reports the aggregate figure of 52% high or critical without specifying the exact split. The two sources converge: the majority of flaws sit in the upper danger band.
By category, 49% are memory corruption, 27% integer issues, 16% input validation, 8% control flow/logic. This technical profile aligns with legacy C/C++ network stacks where manual memory management is standard and bounds checks are scattered.
"AI did not eliminate the need for expert vulnerability researchers, but it dramatically compressed the work" — Daniel dos Santos, VP of Research at Forescout
The Vendor That No Longer Exists
Among Forescout's investigations, one involved firmware for a discontinued wireless access point from a defunct vendor. The system found 7 device-specific vulnerabilities. Among them: unchangeable admin credentials and a buffer overflow in the firmware upload process. No vendor exists to receive reports or issue patches.
This isn't an edge case. It's the core of what Forescout calls the "vulnerability orphanage": devices that stay on networks for operational reasons, with lifecycles far exceeding those of the software governing them. Siemens has identified APOGEE, Desigo, and TALON building automation controllers as affected by the discovered vulnerabilities, confirming the attack surface touches real critical infrastructure.
Daniel dos Santos, VP of Research at Forescout, stated: "Our research demonstrates that AI is changing the economics of vulnerability discovery and accelerating the pace at which security teams need to identify, prioritize, and mitigate cyber risk. This is especially impactful for critical infrastructure, where legacy and difficult-to-patch devices may remain in service for years." The same source reports another dos Santos statement: "With access to an effective AI-assisted workflow, vulnerabilities in complex embedded software can be found much faster and at a fraction of the cost, by security researchers and attackers alike." The "alike" isn't rhetoric: the asymmetry is that the tools are accessible to both sides.
What to Do Now
For security teams managing embedded devices and OT infrastructure, the research demands three concrete actions. First: inventory the TCP/IP stacks in use, with particular attention to IPv6, HTTP/web server, and Telnet components that concentrate 64% of discovered flaws. Second: identify devices from defunct or discontinued vendors and isolate them via network segmentation, since patching is excluded by definition.
Third: evaluate adopting AI-assisted workflows for internal vulnerability research, considering that the $935 cost for static inspection makes a research cycle accessible that previously required months and specialist budgets. The time compression — from nine months to 90 hours — means an offensive actor can iterate more research cycles in the time a defensive team completes one.
The structural limit remains patching. Of 54 vulnerabilities reported to vendors with existing products, only 5 had been fixed at publication. For devices from defunct vendors, fixing is impossible by definition. Barry Mainz, CEO at Forescout, stated: "Organizations still have legacy systems, operational requirements, and devices that can't easily be patched or simply be taken offline. Containing risk, rather than patching, must be the priority." This is a paradigm shift: from remediation to containment.
The impact for critical infrastructure operators is twofold. Acceleration of offensive risk: more flaws found in less time, with costs that don't filter out non-state actors. And the need to recalibrate strategies toward segmentation, visibility, and access control, since patching isn't an option for broad swaths of OT/IoT.
What Remains Undocumented
The dossier does not specify which exact version of Claude was used. Help Net Security separately reported on Claude Opus 5.5, but it is not verified that this model was the one employed in the Forescout study. The names of three TCP/IP stacks were withheld during disclosure for coordination purposes. The specific defunct vendor and discontinued product name are not disclosed. No evidence emerges that the vulnerabilities were already exploited in the wild before discovery.
Frequently Asked Questions
Can all affected devices be patched?
No. The research explicitly documents cases of defunct vendors and discontinued products for which patching is impossible. Even for existing vendors, only 5 of the 54 reported vulnerabilities had been fixed at publication time.
Has AI replaced human researchers?
No. According to Daniel dos Santos' statement reported by Industrial Cyber, AI compressed the work but did not eliminate the need for experts. The 90 total hours include roughly 80 hours of human validation after 10 hours of automated analysis.
Does the $935 include system development?
No. The $935 cost refers to the automated code inspection within the context of an already-completed system build. The estimated total costs of $1,200 to $2,300 include dynamic testing but not the initial investment in building the system.
Sources
Get DeafLetter
A weekly selection of signals, vulnerabilities and guides. Critical alerts remain optional.
You can unsubscribe at any time. Privacy policy.