AI-assisted review finds 99 flaws in widely used IoT TCP/IP stacks

src-7e385673

Forescout says a specialized artificial intelligence harness identified 99 vulnerabilities across 10 TCP/IP software stacks in 90 hours, a review effort the company estimated would have taken nine months without the technology. The findings include flaws in embedded software used in internet of things and industrial devices, underscoring how longstanding weaknesses in foundational networking code can persist across products.

The research, reported by Forescout on Nov. 12, 2026, also illustrates the limits of automated analysis. The harness was effective at finding certain classes of coding errors, particularly memory-safety issues in protocol parsing, but it did not systematically identify logic and design flaws.

Review covered ten protocol stacks

Researchers used a harness built with Claude models to examine 10 popular TCP/IP stacks. Seven had been included in Forescout’s earlier TCP/IP vulnerability research, while three were new to this review. Most of the projects are open source.

The software examined included Contiki, which is used in low-power IoT devices such as street lights; Ethernut, used in Ethernet-based smart devices; lwIP, a lightweight TCP/IP implementation; and wolfIP, used in real-time operating systems for safety-critical devices. These components can be incorporated into products from multiple manufacturers, so a flaw in a shared stack may have implications beyond a single device line.

Of the 99 findings, 33 were vulnerabilities Forescout had previously identified in stacks it had tested. The other 66 were new discoveries across those stacks and the previously untested projects. Forescout estimated the harness’s token costs at $1,200 to $2,300, describing that expense as roughly 100 times less than the salary cost of a senior vulnerability researcher conducting a manual code review.

Memory-safety flaws dominated

About half of the vulnerabilities involved memory corruption. Forescout classified 27% as integer issues, 16% as input-validation failures and 8% as control-flow or logic flaws. The percentages total slightly more than 100%, apparently because of rounding in the reported breakdown.

The researchers said the results reflected the harness’s focus on static-analysis rules for generic code patterns. Such rules are better suited to detecting memory-safety problems than implementation-specific logic errors, which can be harder to recognize without an unacceptable tradeoff between false positives and false negatives. The report said the harness found logic bugs incidentally rather than through a targeted method.

That distinction matters for organizations assessing AI-assisted code review. Faster scanning can expand the amount of code that receives scrutiny, but these results do not show that automated analysis can replace expert review or reliably uncover every class of defect. The findings instead point to a complementary use: applying automated methods to identify patterns at scale, while retaining deeper review for design and logic problems.

Disclosure and patch status

Forescout said it reported 54 of the 66 newly discovered vulnerabilities to the maintainers of the relevant products. The company did not report the other 12 because it believed they were absent from the latest versions of the affected products. Among the reported issues, 24 had been fixed, 29 were still under investigation, and one involved Ethernut, whose developer had not responded to the disclosure.

The status information means that not all reported findings had a confirmed fix at the time of the report. Device makers and operators should therefore avoid assuming that a vulnerability is remediated simply because it has been disclosed. They need to track affected components and versions through vendors and maintainers, and confirm whether updates apply to the products they actually deploy.

Legacy equipment remains a concern

Forescout warned that vulnerable stacks could affect hundreds of products across several industries. A particular concern is discontinued equipment that remains in service and may no longer receive vendor support. The researchers found seven vulnerabilities in a TCP/IP stack used in an unsupported wireless access point; one issue involved hardcoded administrator credentials. Because the access point was unsupported, a vendor patch might not be available.

The review also revealed vulnerabilities in code that Forescout researchers and other experts had examined previously. That experience highlights the challenge of establishing confidence in mature, widely reused networking components: prior audits do not guarantee that a project is free of flaws, and newly discovered issues may have consequences for downstream products. For infrastructure teams, maintaining an inventory of embedded devices, identifying unsupported equipment and checking the status of shared software components are practical steps toward understanding exposure.