AI Cut Vulnerability Research From Nine Months to 90 Hours, and Google Just Cut Your Patch Window Too
Cybersecurity

AI Cut Vulnerability Research From Nine Months to 90 Hours, and Google Just Cut Your Patch Window Too

Forescout found 99 flaws across ten TCP/IP stacks for under 2,300 dollars using off-the-shelf AI models. The same week, Google Project Zero cut its disclosure deadline to 30 days, citing the exact same shift in attacker economics.

PublishedNovember 19, 2026
Read time5 min read
Share

What Forescout actually did

Forescout's Vedere Labs team built a research pipeline combining generally available AI models, including Claude, with existing security tooling to hunt for vulnerabilities in embedded TCP/IP network stacks, the software layer that handles network communication inside everything from industrial controllers to medical devices to ordinary network switches. The pipeline examined both open-source code and compiled software without access to source, a notably harder analysis target, producing 362 initial findings that were narrowed to 312 candidates after deduplication and ultimately 99 confirmed, validated vulnerabilities spread across ten entirely different network stacks from different vendors.

Of those 99, 66 were previously unknown to the security research community, and 52 percent of those new findings carried high or critical severity ratings. The team also rediscovered 33 of 82 known vulnerabilities used as a validation benchmark, confirming the AI-driven approach was not simply flagging noise but actually finding real, previously catalogued flaws alongside genuinely new ones.

The economics that should worry you

The comparison Forescout draws explicitly is to Project Memoria, its own prior TCP/IP stack vulnerability research effort from 2020 to 2021, which took approximately nine months of manual analyst work to complete. This new pipeline reached a comparable result in about 90 hours total: roughly 10 hours of automated analysis plus 80 hours of human validation. Automated code inspection alone cost 935 dollars; factoring in the compute spent running and testing the software under analysis, total model costs landed between 1,200 and 2,300 dollars.

Daniel dos Santos, Forescout's VP of Research, put the implication plainly: this changes the economics of vulnerability discovery, for defenders and attackers alike, since the same generally available models and techniques are not exclusive to security researchers acting in good faith on a defender's payroll. A capability that used to require a nine-month, well-funded research engagement with a dedicated team is now reproducible by a well-resourced attacker for a few thousand dollars and a handful of days, a gap that should reshape how any security leader thinks about the realistic shelf life of an unpatched system in their own environment.

The problem with no patch at the end of it

A genuinely uncomfortable finding in this research involves devices that have outlived their manufacturer entirely. Forescout's example is a discontinued wireless access point where the pipeline found seven device-specific flaws, including administrator credentials that could not be changed and a buffer overflow in the firmware upload process, with no vendor remaining to issue a fix for any of it. Siemens separately confirmed affected building automation controllers across its APOGEE, Desigo, and TALON product lines, products still deployed in active commercial buildings.

Barry Mainz, Forescout's CEO, framed the practical response bluntly: containing risk, rather than patching, must be the priority for this category of device. That is a genuine shift in posture for security teams accustomed to a patch-first mental model. When AI-accelerated research finds flaws faster than vendors can or will fix them, isolating exposure, through network segmentation, access restriction, and monitoring, becomes the primary control rather than the stopgap while a patch is pending.

Why Google moved its disclosure deadline the same week

Google Project Zero announced a new disclosure policy, which it calls Full X, cutting its standard vulnerability disclosure deadline from the prior 90-plus-30-day model down to a flat 90 days immediately, stepping down to 60 days on February 2, 2027, and 30 days on January 3, 2028. The team's own explanation ties directly to the same dynamic Forescout's research demonstrates: large language models are shortening the time from patch availability to working exploit, in some reporting to as little as 30 minutes, which makes the old grace period after a patch shipped largely meaningless as protection.

Project Zero's Natalie Silvanovich put the reasoning in terms every security leader should sit with: a vulnerability reported today presents a much higher immediate risk to users than one reported a year ago, because the tools available to turn a disclosure into an exploit have gotten dramatically faster and cheaper in that same period. Project Zero now also finds several vulnerabilities a week, up from a few per year historically, making its old negotiation-heavy disclosure process increasingly unworkable at the new discovery volume.

What this means for your own patch cycle

Two independent developments in the same week point at the same structural shift: AI is compressing both ends of the vulnerability lifecycle, discovery and exploitation, at the same time vendors are compressing the disclosure window in response. The practical effect on your organization is that the gap between a flaw becoming public and a working exploit existing against it is shrinking from months to potentially minutes, and your patch management process needs to assume that compression rather than the multi-week runway it may have been built around.

Audit your own patch SLAs against this new reality specifically, not against the threat model your process was originally designed around last year or the year before. For any system where patching is slow, discontinued, or simply not possible given the vendor's own lifecycle decisions, start building the containment-first posture Forescout's CEO described now, before an AI-assisted researcher or attacker finds the next flaw in it, rather than scrambling to react after the fact. The nine-month research project is not coming back as a reasonable planning assumption, and your organization's patch and disclosure planning window should stop assuming it will.

Tagged#news#security#cybersecurity#breach#cisa#ransomware#zero-day#supply-chain#ai-security#forescout#vedere-labs#google-project-zero#tcp-ip-stacks#disclosure-policy#ai-assisted-research#embedded-devices