Forescout researchers used AI models and security tools to identify 99 vulnerabilities across ten TCP/IP stacks, including 66 previously unknown flaws. Help Net Security reports that 52% of those newly discovered vulnerabilities were rated high or critical, a striking result in software that can remain inside connected devices for years.
Watch Desk analysis
What happened
The researchers built a system around Claude models, existing security tools and 32 scanning rules. The system first flagged suspicious code patterns, then used AI agents to investigate whether an attacker could reach and exploit the suspected weakness. Researchers could intervene and guide the process.
The reported 99 confirmed vulnerabilities emerged from 362 initial findings, narrowed to 312 candidates after duplicates were removed. Forescout said several projects have issued fixes for some of the flaws; investigations into three networking stacks remain open, with their names withheld during disclosure.
Help Net Security’s report says the researchers’ automated analysis took about ten hours, followed by roughly 80 hours of validation. They compared that work with around nine months of manual code analysis in an earlier investigation. The figures describe the completed system applied to these targets, not the additional effort needed to build and refine it.
Key findings
- 99 vulnerabilities across ten stacks
Forescout’s researchers confirmed the flaws in software used by connected products. - 66 flaws were previously unknown
Of those newly discovered vulnerabilities, 52% were rated high or critical. - Testing filtered a large first batch
The system produced 362 initial findings; testing and review confirmed 99 vulnerabilities. - Compiled software took more work
The report says analysis without source code cost about six times as much and took roughly twice as long as source-code inspection.
Why it matters
The practical significance is not simply that AI can inspect code faster. The research points to a way of helping security teams find weaknesses in software that may be difficult to examine, while also showing why automated findings need extensive validation. Here, testing and review did much of the work of separating real vulnerabilities from the initial pile of leads.
Some affected devices may be hard to patch, and at least one product came from a company that had ceased operating. That leaves a stubborn security problem: finding a flaw does not conjure a vendor able to fix it. For organisations running legacy equipment, containment may be the only available response, as Forescout CEO Barry Mainz told Help Net Security.
Our read
This is a substantial result, and the best detail is the gap between ten hours of automated analysis and 80 hours of validation. AI appears to have made vulnerability discovery more accessible in this study, but the researchers’ own figures make clear that a promising automated finding is not yet a confirmed security flaw.
The work also has limits: the system rediscovered 33 of 82 known vulnerabilities used for comparison, and it was better at finding memory errors than logic or design problems. Treat it as a useful research result, not a promise that AI can sweep every network stack clean.
What to watch
- Whether more maintainers confirm fixes, especially for products with no active vendor.
- How the researchers’ method performs on additional software and vulnerability types.
- Whether security teams can validate AI-generated findings quickly enough to keep pace with new discoveries.
Discussion spark: If AI makes it cheaper to uncover flaws in old network software, who should carry the responsibility when the vendor no longer exists: device owners, researchers, or the wider industry?
Sources and evidence
- Source update (16 November 2026, 05:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.