Forescout Reports 66 New TCP/IP Flaws Found With AI

Forescout’s Vedere Labs says an off-the-shelf AI system, pointed at ten TCP/IP stacks used in connected industrial and IoT equipment, found 99 TCP/IP flaws, 66 of them new, in about 90 hours of work. The company published the research and a press release on 12 November 2026. It compares that time with about nine months of earlier manual analysis, a comparison that is not a controlled, like-for-like benchmark, and it gives a static-analysis cost of about $935 that excludes the work of building the testing harness.

The research report, “Finding Critical Infrastructure Bugs with Off-the-Shelf AI,” says the team built a specialised harness for embedded software and ran it against ten popular TCP/IP stacks: seven it had previously analysed as part of a set it calls Project Memoria+, and three new ones, lwIP v2.2.1, wolfIP v1.0 and a third, labelled S3, whose vendor is still in a disclosure process. Of the 66 new findings, 54 were reported to vendors and maintainers: 20 in open-source stacks and 34 in proprietary stacks. The research page says that, of those proprietary cases, “only five have been fixed to date.” The press release says “52% of the 66 new vulnerabilities were rated high or critical severity.” (The research page’s shorter wording, “52% of the vulnerabilities,” is broader; the press-release phrasing is the one used here.)

The time and cost figures, and what they leave out

Forescout says the 90 hours were made up of roughly ten hours of independent AI analysis and about 80 hours of dynamic validation, and that the system rediscovered 40 percent of vulnerabilities the lab had previously disclosed on the stacks it had already studied. The press release presents this as “dramatically reducing the time and cost required for security researchers – and attackers alike – to find vulnerabilities.” Daniel dos Santos, Forescout’s vice-president of research, is quoted as saying that AI “did not eliminate the need for expert vulnerability researchers, but it dramatically compressed the work.”

Two limits sit underneath those headlines. First, the nine-month comparison is with the lab’s previous analysis, not a controlled benchmark on the same targets under the same rules. The new run includes three stacks that were not in the earlier set, and the older work was a different project. Second, the report is explicit that the cost comparison “does not include the effort to design the AI harness, develop, and refine static analysis rules, build the emulation infrastructure, and refine prompts until the agents behaved consistently.” The $935, or $2.58 per finding, is described as the marginal cost of applying the finished harness to new targets. Binary analysis, the lab says, was on average six times more expensive and took twice as long as source-code analysis.

The models were from the Claude family. Forescout says it chose them for earlier evaluation results and for “not-too-restrictive guardrails,” using Sonnet 4.6 with high and max reasoning in the first stages and Opus versions between 4.6 and 5 for dynamic verification.

Siemens has patched some products and listed others as having no fix

One official downstream advisory already on the record is Siemens ProductCERT SSA-333035, published on 13 October 2026 and written in coordination with this research. Siemens gives the advisory a CVSS v3.1 base score of 8.2 and a CVSS v4.0 base score of 8.8. It has issued updates for APOGEE PXC Compact and Modular controllers: V3.5.7 for the BACnet versions and V2.8.21 for the P2 Ethernet versions. For APOGEE MBC and MEC, in all listed versions, and for Desigo PXC and PXM products at version 2.3 or later, Siemens says “Currently no fix is planned.”

The mitigations Siemens lists are practical rather than a software patch: put the controllers behind network segmentation and properly configured gateways or firewalls, and, where relevant, leave the FTP and web services turned off. Siemens notes that the FTP and web services are disabled by default on the APOGEE, Desigo and TALON products. CISA advisory ICSA-26-295-08, released on 22 October 2026, repeats that Siemens has new versions for several products and recommends those mitigations where a fix is not available. It lists affected APOGEE, TALON and Desigo PXC/PXM products.

Forescout’s own advice to asset owners is in the same vein: restrict unnecessary services, isolate devices that cannot be patched, and shorten the time to apply fixes when they exist, on the grounds that the pace of disclosure may now be faster. The lab also warns that stack-level patches still have to reach devices through the supply chain, and that legacy equipment may be difficult or impossible to patch.

That supply-chain lag is familiar from other embedded and industrial flaws TSN has covered, including GoAhead’s TLS fixes for an embedded web server, a Siemens Software Center flaw that was exploited to infect Windows PCs, and a CISA KEV order that included a Siemens product among systems agencies were told to patch.

A faster hunt is not the same as a closed case

The useful claim in the Forescout paper is not that nine months of work has been scientifically shown to shrink to 90 hours on identical targets. It is that a generally available model, wrapped in a harness the lab had to build first, produced 66 new TCP/IP flaws, plus a 40 percent rediscovery rate on previously studied stacks, at a modest marginal cost. The nine-month figure is a comparison with earlier work, and the $935 figure starts after the harness exists. Anyone repeating the exercise would still have to pay, in time or money, for that upfront build.

The more immediate question is how much of the new list gets fixed. Fifty-four of the 66 findings were reported; only five of the 34 proprietary cases had been fixed when the research was published, and Siemens has already drawn a line under some older APOGEE and Desigo controllers, offering segmentation and default-off services rather than a new firmware release. For operators, the defensive steps are the ones already in the Siemens and CISA notices: segment the controllers, keep FTP and web services off unless they are required, and apply V3.5.7 or V2.8.21 where those updates exist. The research is a warning about the pace at which flaws can now be found, including by attackers. It is not itself an incident.