AI finds zero-day flaws in the networking software inside critical infrastructure at $14.17 a flaw, and most of the flaws it found are still unfixed. Forescout Research’s Vedere Labs used off-the-shelf AI models to find 66 previously unknown vulnerabilities in 10 popular TCP/IP stacks in 90 hours, at a static-analysis cost of $935. Just over half of the 66 new flaws, 52%, were rated high or critical severity. Only 24 of the 66 were fixed when the report was published on 12 November 2026. No exploitation in the wild is reported, and no source claims any of the 66 flaws is being exploited. This is a research and disclosure story, not an incident. The headline that a zero-day now costs $4 is wrong: $4.16 is the top of the per-candidate-finding range, and the honest per-flaw cost is $14.17 for static analysis, or $18.18 to $34.85 once all model spending is counted. That is still trivially cheap, which is the point.
What the AI hunt actually found
Forescout’s team built a specialised AI-assisted vulnerability research harness and ran it against 10 popular TCP/IP stacks: seven previously analysed in Project Memoria in 2020 and 2021, plus three new ones. The harness found 99 vulnerabilities in 90 hours: 33 previously known vulnerabilities rediscovered, 40% of the 82 in what the report calls “Project Memoria+”, and 66 previously unknown vulnerabilities. The stacks tested were Contiki v3.0, Ethernut v5.1, FNET v4.6.3, PicoTCP v1.7.0, Nucleus NET v5.1 from Siemens, lwIP v2.2.1 and wolfIP v1.0. Three proprietary stacks are anonymised as S1, S2 and S3 because responsible disclosure is still ongoing, and their vendors and products are not identified.
Severity is Forescout’s own rating of its findings, not a vendor-confirmed or CVE-scored rating, and no CVE IDs were cited for the 66 in the sources read for this article. Even so, 52% of the 66 new vulnerabilities were rated high or critical severity, and 29% affect core components of the TCP/IP stacks rather than application-level or other protocols. Disclosure is under way: 54 of the 66 new vulnerabilities were reported to the respective vendors or maintainers, 20 affecting open-source stacks and 34 affecting proprietary stacks, and coordinated disclosure continues for the remainder. That leaves 12 of the 66 not yet reported to any vendor. The report says 54 were reported “because we believed they still affected latest versions of their products.” The report does not state the reason for each of the other 12, so this article does not claim the other 12 do not affect latest versions.
The fix status at publication, from the report’s own status table, is blunt. FIXED: Contiki 4, FNET 3, PicoTCP 5, Nucleus NET 5, wolfIP 3, lwIP 4, a total of 24 of 66. NO RESPONSE: Ethernut 1. INVESTIGATING: S1 7, S2 10, S3 12, a total of 29 of 66. The remaining 12 were not yet reported. Forescout’s supply-chain warning is that “past disclosures suggest these supply-chain flaws likely affect hundreds of products from dozens of vendors. However, stack vendors’ patches must still reach affected devices through the supply chain.” Project Memoria previously identified more than 100 affected products from dozens of vendors. That “hundreds of products” figure is Forescout’s estimate based on past disclosures, not a counted list for these 66 flaws, and no user count exists.
The fix gap is the story
Only 24 of 66 flaws, 36%, were fixed at publication. Another 29, 44%, are still under investigation with the vendors unnamed. One drew no vendor response at all, and 12 were not yet reported. Nearly two-thirds of the flaws this cheap AI hunt found are still live in the supply chain. That is the point of Forescout CEO Barry Mainz’s line that security teams “simply cannot patch at the speed we’re now seeing vulnerabilities being discovered.”
Siemens was first to identify vulnerable devices: several building-automation controllers in the APOGEE, Desigo and TALON product lines run Nucleus NET and are susceptible to the newly discovered issues. One case is the unpatchable legacy exposure. One analysed stack’s vendor general support ended in December 2025, with extended support ending September 2027. The S1 stack was embedded in the firmware of a discontinued wireless access point from a company that is now defunct, and the harness found 7 vulnerabilities specific to those devices, including buffer overflow via firmware upload and hardcoded administrator credentials that cannot be changed for web and shell management. No maker remains to issue a firmware fix for those discontinued devices.
How AI finds zero-day flaws
The harness runs Claude-family models, Sonnet 4.6 for the first three stages and Opus 4.6 to 5 for dynamic verification, alongside deterministic static analysis with Semgrep and Joern, 32 Joern rules for memory corruption and protocol parsing, and Ghidra to convert binaries to a common intermediate representation. Dynamic verification generates proof-of-concept scripts to confirm each finding. The time comparison is the headline number: Project Memoria’s manual code analysis took about nine months, while the AI-assisted approach took 90 hours total, about 10 hours of independent AI analysis plus roughly 80 hours of human validation and dynamic verification. On a straight calendar comparison, about 270 days times 24 hours is about 6,480 hours against 90, roughly 72 times faster. On working hours, nine months at about 160 hours a month is about 1,440 hours against 90, roughly 16 times. TechReview analysis: either way, the bottleneck has moved from finding bugs to fixing them.
That speed comes with a precision reality check. The harness produced 362 static-analysis findings, which became 312 unique candidates after deduplication, then 99 confirmed issues. Only 99 of 312 unique candidates were confirmed, 31.7% overall, and average per-stack precision was 42%, ranging from 11% on the worst stack, S1 with 73 candidates and 8 confirmed, to 70% on the best, S2 with 23 candidates and 16 confirmed. Source-code targets averaged 47% precision against 34% for binaries, 22% excluding the S2 outlier, whose demo binary retained symbols. Roughly two in three AI-flagged candidates were false leads, and about 80 of the 90 hours went to human validation. The “AI replaces researchers” framing is wrong. As Forescout VP of Research Daniel dos Santos put it, “AI did not eliminate the need for expert vulnerability researchers, but it dramatically compressed the work.”
Binary targets were also far more expensive to analyse than source code: static analysis of binary targets was on average six times more expensive and took twice as long, with source-code targets averaging under 40 minutes each against around 90 minutes for binary targets. Dynamic analysis was estimated at $15 to $30 per open-source target against $50 to $300 per binary target.
The $4 headline is wrong: here is the real cost
ZDNet’s headline, “AI finds zero-day flaws for just $4 now – even mediocre hackers can cash in”, published 16 November 2026. The $4 figure is not directly verified against ZDNet’s body text, which could not be read for this article, but it is consistent with the primary source in one narrow sense: Forescout’s own per-finding static-analysis cost ranged from $0.95 to $4.16, topping out just above $4, with an average of $2.58 per finding across 362 candidate findings processed. The $4 only works as the top of that per-candidate-finding range. It is not a per-zero-day price, and this article does not present it as one.
The corrected per-flaw maths, from the report’s own numbers, is $935 of static analysis divided by 66 new flaws, $14.17 per previously unknown flaw. Total model spend of $1,200 to $2,300 divided by 66 is $18.18 to $34.85 per new flaw. Even the most expensive reading stays under $35, but it is not $4. A headline that says a zero-day costs $4 understates the real per-flaw cost by about 3.5 times, and the real cost is still trivially cheap. That is the honest version of the story. The reported cost is also the marginal cost of applying the completed harness to new targets. It does not include the effort to design and build the harness, develop static-analysis rules, build emulation infrastructure and refine prompts. Forescout says the cost is likely 100 times less than the cost of a senior vulnerability researcher doing manual code analysis.
The “even mediocre hackers can cash in” framing is headline only. The Forescout report tests whether attackers could use generally available models. It reports no real attacker using this approach, no exploitation in the wild, and no bounty payouts or “cashing in”. None of that is presented as fact here.
Finding bugs is nearly free. Weaponising and patching them is not.
The comparison that puts Forescout’s numbers in context is OpenAI’s Linux study from November 2026, where an autonomous pipeline found and wrote working exploits for 948 Linux privilege-escalation bugs in 30 hours for about $1.2 million, about $1,279 per functional proof of concept, including exploit development and draft patches. Forescout found 66 new flaws in 90 hours at $2.58 per candidate finding, with no working exploit produced. The roughly 500-times gap between $2.58 and $1,279 is the price of turning a finding into a working exploit. TechReview analysis: finding the bug is now nearly free; weaponising it is where the money goes. The two studies have different targets, different outputs and different cost bases: Forescout tested TCP/IP stacks and produced candidate findings, while OpenAI tested the Linux kernel and produced working exploits plus draft fixes.
Forescout cites PHP2Shell as one of this year’s biggest examples of AI-discovered vulnerabilities. TechReview covered the PHP2Shell disclosure, when PHP shipped emergency fixes on 21 October 2026 for two critical flaws, CVE-2026-109618 and CVE-2026-109619, that let an attacker run code of their choosing on a vulnerable server without logging in. The 66 flaws in this study are in TCP/IP stacks, not PHP, and PHP2Shell is not a finding of this study.
For the devices around you, Forescout’s recommendations are practical. Asset owners should prioritise by reachability, exploitability and operational impact, reduce device exposure by restricting FTP, Telnet, IPv6 and web administration interfaces, isolate devices that cannot be patched, and shorten patching timelines. Vendors should use AI to find bugs before attackers do and prepare for more AI-generated vulnerability reports.
What happens next
Four things would change this story. Disclosure of the S1, S2 and S3 stack identities and their vendors’ fixes. Siemens advisories for the APOGEE, Desigo and TALON controllers running Nucleus NET. Any response from Ethernut, the one vendor that had not responded at publication. And CVE assignments for the 66 new flaws, none of which were cited in the sources read for this article. Whether ZDNet’s “$4” framing gets echoed or corrected by other outlets is worth watching too. Until then, the confirmed facts are narrow: off-the-shelf AI found 66 new flaws in 90 hours for $935 of static analysis, 52% of them high or critical, and only 24 were fixed at publication on 12 November 2026.


