AI Slashes Vulnerability Discovery Time in Critical Infrastructure

The cybersecurity landscape has reached a structural inflection point. Forescout Research’s Vedere Labs published findings on November 12, 2026, demonstrating that off-the-shelf artificial intelligence models can compress vulnerability discovery in foundational networking software from approximately nine months to just 90 hours. The research, titled “Finding Critical Infrastructure Bugs with Off-the-Shelf AI,” uncovered 99 vulnerabilities across 10 widely used TCP/IP stacks, including 66 that were previously unknown. More than half of the new discoveries were rated high or critical severity.

From Nine Months to 90 Hours

The numbers alone tell a staggering story. Forescout built a specialized AI-assisted vulnerability research harness using generally available models and applied it to 10 popular TCP/IP stacks, the foundational software that enables connected devices to communicate over networks. Seven of these stacks had been previously analyzed under Project Memoria, an initiative that began in 2020, took 18 months, and found nearly 100 CVEs across 14 embedded TCP/IP stacks. Three were entirely new targets.

The AI harness processed all 10 stacks in approximately 90 hours of wall-clock time. Source-code targets averaged under 40 minutes each, while binary targets took roughly 90 minutes, about 2.3 times longer. The total cost for static analysis across all stacks was approximately $935, averaging $2.58 per finding. By comparison, the original Project Memoria required seasoned researchers working manually for a year and a half to achieve comparable scope.

This is not a marginal improvement. It is a fundamental reordering of the economics of vulnerability research. Tasks that once required deep expertise, significant time, and substantial budgets can now be executed by a single researcher with AI assistance at a fraction of the cost.

What the AI Found

Out of 362 candidate findings generated by static analysis, the system validated 99 confirmed vulnerabilities after deduplication and AI-driven reasoning. Thirty-three of these were previously known vulnerabilities that the AI rediscovered, demonstrating a 40 percent rediscovery rate. The remaining 66 were entirely new discoveries.

The severity breakdown is alarming:

  • 47 percent of new vulnerabilities were rated high severity (CVSS 7.0 to 8.9)
  • 5 percent were rated critical (CVSS 9.0 or higher)
  • 45 percent were medium severity (CVSS 4.0 to 6.9)
  • 3 percent were low severity (CVSS below 4.0)

Memory corruption bugs accounted for the largest share at 49 percent, covering out-of-bounds reads and writes and buffer overflows. Integer issues followed at 27 percent, including underflow, overflow, truncation, and sign errors. Input validation flaws comprised 16 percent, while control flow and logic bugs made up the remaining 8 percent.

By affected component, core networking protocols (IP, TCP, UDP, ICMP, and Ethernet) accounted for 29 percent of new findings. HTTP and web servers represented 15 percent, while IPv6, Neighbor Discovery, RPL, and DHCP clients each accounted for 10 percent. Telnet comprised 7 percent, and DNS and management command-line interfaces each accounted for 2 percent.

The Supply Chain Ripple Effect

TCP/IP stacks are not standalone applications. They are embedded in products across operational technology, Internet of Things devices, medical equipment, network infrastructure, and other cyber-physical systems. A single vulnerability in a shared TCP/IP implementation can propagate through hundreds of downstream products from multiple manufacturers.

Forescout identified Siemens as the first vendor with confirmed affected devices. Several building automation controllers in its APOGEE, Desigo, and TALON product lines run Nucleus NET and are susceptible to the newly discovered issues. Other vendors are still investigating their exposure.

The supply chain problem compounds when vulnerable software is embedded in devices that remain in service for decades. Even when a patch is available, it must travel from the software developer through firmware and product vendors before reaching the organization operating the affected device. For discontinued or unsupported products, a patch may never arrive at all. In one case, Forescout analyzed a TCP/IP stack embedded in the firmware of a discontinued wireless access point made by a now-defunct company. The AI harness identified seven vulnerabilities specific to these devices, including a buffer overflow via firmware upload and hardcoded administrator credentials that cannot be changed.

Binary Analysis Opens New Attack Surfaces

One of the most significant aspects of the research is the AI harness’s ability to analyze proprietary binaries where source code is unavailable. Binary targets were more complex and expensive to process, with static analysis averaging six times the cost and taking twice as long compared to open-source targets. Decompiled targets produced two to three times more candidate findings than open-source stacks, and five to ten times more survived reachability filtering.

Inaccuracies from the Ghidra decompiler introduced additional false positives, each of which can raise the cost per finding by three to seven times if not caught early. Despite these challenges, the harness successfully identified real vulnerabilities in binary-only targets. This capability matters enormously for critical infrastructure, where proprietary firmware is the norm and source code is rarely available to external researchers or attackers.

Implications for Defenders

The Forescout research carries several critical implications for security teams and infrastructure operators:

  • Patching alone is no longer viable. The accelerated pace of discovery means vendors must address a growing volume of patches, and end users cannot rely on patching as their primary defense. Many OT devices cannot be patched quickly, and some will never receive a fix.
  • Containment must replace remediation as the priority. Organizations need to know what is connected, understand which vulnerabilities create real operational risk, and maintain controls that contain that risk rather than waiting for patches that may never come.
  • End-of-life devices require isolation. Unsupported products should reside in tightly restricted network segments that allow only essential traffic, with replacement plans in place. AI makes it inexpensive to find new flaws in these devices, but no vendor will ever fix them.
  • Security must become part of procurement. Suppliers should be required to state support lifetimes, publish vulnerability disclosure policies, provide software bills of materials, and demonstrate evidence of ongoing security testing including AI-assisted testing.
  • Internet-facing management interfaces demand immediate attention. Exposed interfaces on vulnerable devices provide the easiest path for exploitation and should be the first targets for access restrictions.

The Broader Trend

The Forescout findings align with a broader pattern documented across the cybersecurity industry in 2026. Google’s Threat Intelligence Group reported that threat actors are now using multi-agent AI frameworks to fully automate vulnerability discovery and exploitation without human intervention for the first time. Palo Alto Networks’ Unit 42 observed that their frontier AI systems discovered 26 CVEs in a single month, compared to their typical volume of fewer than five. Mandiant’s M-Trends 2026 report documented that the mean time to exploit vulnerabilities has turned negative, averaging negative seven days, meaning exploitation now routinely precedes public patch availability.

The Forescout research adds a critical dimension to this picture. It demonstrates that the barrier to entry for sophisticated vulnerability research has dropped dramatically. A single researcher with approximately $935 in AI API costs and 90 hours of processing time can now replicate work that previously required a team of experts working for over a year. This capability is available to defenders and attackers alike.

What Organizations Should Do Now

Security teams should treat the Forescout findings as a planning catalyst rather than a theoretical concern. Immediate actions include:

  • Inventory all embedded devices running TCP/IP stacks, including OT controllers, IoT sensors, medical devices, and network equipment. Document firmware versions and vendor support status.
  • Engage vendors proactively to ask whether their products use any of the analyzed TCP/IP stacks and whether patches are planned for the 66 newly reported vulnerabilities.
  • Implement network segmentation for devices that cannot be patched, isolating them from general corporate networks and the internet where possible.
  • Deploy secure remote access solutions that replace flat VPN access and shared credentials with brokered, least-privilege access, reducing the attack paths available to adversaries.
  • Monitor for exploitation indicators and track security advisories as vendors disclose and patch the new vulnerabilities over coming months.

The message from Forescout’s research is unambiguous. AI has fundamentally changed the economics of vulnerability discovery. Organizations that continue to treat patching as their primary security strategy will find themselves increasingly exposed as the gap between discovery and remediation continues to widen. The future of critical infrastructure security lies in comprehensive asset visibility, aggressive containment of unpatchable systems, and procurement practices that demand security accountability from every supplier in the chain.


Edited by Palawan @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous


Discover more from QUE.com

Subscribe to get the latest posts sent to your email.

Leave a Reply

Connect with

Discover more from QUE.com

Subscribe now to keep reading and get access to the full archive.

Continue reading