Forescout AI harness finds 66 new flaws in TCP/IP stacks used in OT and IoT devices
Forescout says an AI harness confirmed 99 bugs in 10 embedded TCP/IP stacks, 66 of them new, in 90 hours for about $935. Some devices get no patch.
By Berkant A.Published
Tech AI LatestForescout's Vedere Labs said on 12 November 2026 that an AI harness built on commercially available models confirmed 99 vulnerabilities in 10 TCP/IP stacks used in industrial, IoT, medical and network devices. 66 were previously unknown and about half of those are rated high or critical. The work took about 90 hours instead of an estimated nine months and cost about $935 in automated analysis. Some affected products are discontinued and will get no fix.
Key facts
- Announced: 12 November 2026 (Forescout Vedere Labs)
- Result: 99 confirmed vulnerabilities in 10 TCP/IP stacks; 66 new, 33 rediscovered from a set of 82 known bugs
- Effort: about 90 hours versus an estimated nine months of manual analysis; automated analysis cost about $935, or $2.58 per finding
- Severity: 52% of the new vulnerabilities rated high or critical, Forescout says
- Disclosure: 54 of the 66 new findings reported to vendors and maintainers; some discontinued products have no planned fix
- Models: Claude models, according to Cybersecurity Dive; Forescout describes them as off-the-shelf commercial models
What Forescout found
Forescout, a company that secures connected devices, published research from its Vedere Labs unit on 12 November 2026 titled Finding Critical Infrastructure Bugs with Off-the-Shelf AI. The team built a harness around generally available AI models and pointed it at 10 TCP/IP stacks, the networking code embedded in industrial controllers, IoT devices, medical equipment and network gear. Seven of the stacks come from Forescout's earlier Project Memoria research and three are new to this study, according to coverage by Help Net Security and Industrial Cyber.
The harness produced 362 initial findings, which fell to 312 after duplicates were removed. Researchers then confirmed 99 real vulnerabilities by testing. Of those, 66 were previously unknown and 33 were rediscovered from a reference set of 82 known flaws, a recall rate of about 40%. Forescout says 52% of the new vulnerabilities are rated high or critical severity. Cybersecurity Dive reported that the harness used Anthropic's Claude models; Forescout's own summary refers to commercially available models without naming a version.
Nine months of work in 90 hours
The headline claim is speed. Forescout compares the study with Project Memoria, its manual review of TCP/IP stacks in 2020 and 2021, and says the new work took about 90 hours, where it estimates manual analysis would have taken about nine months. The automated analysis cost about $935, or roughly $2.58 per finding, according to Forescout's figures as reported by Infosecurity Magazine and others. Those costs cover running the models, not the time of the researchers who built the harness and checked every result by hand.
The study also shows the limits. The harness missed about 60% of the known reference bugs, and only about a third of its deduplicated findings survived testing. That pattern matches what other labs report: Microsoft needed 26,502 AI findings to produce 59 Linux kernel bug reports, and our comparison of AI bug hunting at four labs found validation, not discovery, to be the bottleneck. What is new here is the target: small embedded codebases in long-lived devices, where the earlier manual effort was slow and expensive.
Devices that will not be fixed
Forescout says it reported 54 of the 66 new findings to vendors and maintainers. Some CVE identifiers for open-source stacks were still being assigned at publication, three proprietary stacks remain under investigation with their names withheld, and Forescout lists the Ethernut project as not having responded. Help Net Security reports that Siemens has published updates for some affected building automation controllers, while several legacy products listed as affected have no planned fix and need network mitigations instead. One stack was found in an unsupported wireless access point that also contained hardcoded administrator credentials.
That is the practical problem for owners of industrial and building systems. An AI harness that costs a few hundred dollars to run can be used by attackers as well as defenders, and discontinued devices have no vendor left to ship a patch. Operators can segment such devices from other networks, monitor their traffic and plan replacement. Google Project Zero has cited the same shift in cutting its disclosure deadline to 30 days by 2028. More coverage is on the cybersecurity hub.
What changed
| Before | Forescout's manual Project Memoria review of TCP/IP stacks took about nine months of analysis |
|---|---|
| After | An AI harness confirmed 99 bugs in 10 stacks, 66 new, in about 90 hours for about $935 in model costs |
| Who it affects | Owners of OT, IoT, medical and network devices built on embedded TCP/IP stacks, and the vendors that maintain them |
Questions people ask
- How many vulnerabilities did Forescout's AI find?
- Forescout says its harness confirmed 99 vulnerabilities in 10 TCP/IP stacks, of which 66 were previously unknown and 33 were rediscovered known bugs.
- Which AI models did Forescout use?
- Forescout describes them as off-the-shelf commercial models. Cybersecurity Dive reported that the harness used Anthropic's Claude models; no specific version is named in the coverage.
- Will the affected devices be patched?
- Not all of them. Forescout reported 54 new findings to vendors, Siemens has updates for some controllers, but several discontinued products have no planned fix and need network mitigations.
Sources
- Finding Critical Infrastructure Bugs with Off-the-Shelf AI (Forescout Vedere Labs, 12 Nov 2026)
- AI uncovers flaws in discontinued devices with no vendor left to patch them (Help Net Security, 16 Nov 2026)
- AI found widespread IoT vulnerabilities more quickly and cheaply than humans could (Cybersecurity Dive, 12 Nov 2026)
- AI Cuts Vulnerability Discovery from Nine Months to 90 Hours (Infosecurity Magazine, 12 Nov 2026)
- AI tears through legacy TCP/IP stacks, exposing vulnerabilities across critical infrastructure, OT and IoT devices (Industrial Cyber, 13 Nov 2026)
Reference pages
Spotted an error? See the corrections policy and tell the editor. Facts in this story come from the sources listed above; vendor claims are attributed.
Comments
Loading comments.





