By Byron V. Acohido
Security teams are being told they have hours to patch.
Related: AI agents have a Lord of the Flies problem
The warning has a real basis. In June, Anthropic’s frontier red team demonstrated an AI model building working exploits from freshly patched Firefox and Windows kernel vulnerabilities, one of them in 31 minutes. Meanwhile ZeroDayClock, widely cited across the security industry, put mean time-to-exploit near a single day and projected it falling to one hour by 2027.
Root Evidence went looking for evidence that attacker behavior had actually changed.
Its recently released Vulnpocalypse Report traced confirmed in-the-wild exploitation of published CVEs from January 2018 through July 15, 2026. Of 253,912 vulnerabilities published during that period, adversaries exploited 3,769, about 1.5 percent. For vulnerabilities exploited after a patch was available, the median defender window was 116 days so far in 2026, compared with 24 days in 2018.
Speed still matters for some flaws. Roughly a third of this year’s exploited n-days were hit within 30 days of a patch.
The report takes AI’s potential seriously. Its focus is the gap between what AI makes technically possible and what attackers are actually doing.
Last Watchdog asked Root Evidence CEO Jeremiah Grossman, who led the analysis, what the data says and what would convince him that AI has changed the game. Here are excerpts of that conversation edited for clarity and length.
LW: What did the record show?
Grossman: Vulnerability exploitation remains one of cybersecurity’s biggest unsolved problems. It accounts for roughly a third of cyber-insurance claims and billions of dollars in losses each year.
Yet after 30 years of chasing vulnerabilities, the record shows attackers need to exploit only a painfully small fraction of them to cause that damage.
That tells us we’ve been fixing the wrong vulnerabilities, in the wrong order. We need to focus on fixing the ones that actually matter.
The mistake we keep making is confusing what can be hacked with what is likely to be hacked.
LW: Why do other measurements show exploitation accelerating?
Grossman: A major reason is where you start the clock. ZeroDayClock measures time-to-exploit from publication in the National Vulnerability Database. We measure from the date a vendor makes a patch available, because that is when defenders can actually act. NVD has recently lagged vendor patches by about 25 days, so measuring from NVD publication can make attackers appear faster even when their behavior has not changed.
LW: What else makes recent years look faster than they are?
Grossman: There is also an aging problem. Imagine watching a marathon two hours after the start. You average the times of the first eight runners across the finish line and conclude the typical marathon takes just over two hours. Thousands of runners are still on the course.
Recent CVE years work the same way. Fast exploitation shows up first. Vulnerabilities exploited months or years later have not yet entered the record. As those later cases are cataloged, the average time-to-exploitation rises. Recent years can look faster than fully aged years even when adversary behavior hasn’t changed.
LW: If AI can build exploits in minutes, why hasn’t attacker behavior changed?
Grossman: Long before generative AI, defenders were already buried in vulnerabilities, unpatched systems and assets they did not know they owned. Yet attackers still exploited only 1 to 2 percent of published CVEs.
Were breaches really constrained by too few vulnerabilities, exploits too difficult to build, or defenders patching too quickly? The argument that AI will dramatically increase breaches assumes those were the constraints in the first place. Maybe vulnerability discovery and exploit development were never the real bottlenecks.
AI does appear to be accelerating vulnerability discovery, and researchers are publishing CVEs faster. What we have not yet seen is a corresponding surge in the share attackers actually exploit, a sustained increase in zero-days or a broad collapse in the defender window.
LW: What would tell you that has changed?
Grossman: I closely monitor breach frequency and financial impact, how attacks are carried out, who is behind them and who the victims are. If we start seeing sustained increases in breach volume or impact, faster exploitation across more and different CVEs, more zero-days, or attacks succeeding at scales that weren’t practical before, that would signal an evolving threat landscape, and AI could be one of the causes.
The important thing is to look for those effects in real-world outcomes, not assume them from what AI is technically capable of doing.
Acohido
Pulitzer Prize-winning business journalist Byron V. Acohido is dedicated to fostering public awareness about how to make the Internet as private and secure as it ought to be.
(Editor’s note: This journalist-led report was produced with underwriting support from some of the featured companies, while Last Watchdog retained full editorial control. I use AI as a tool to help me gather evidence, test ideas and draft. I supply the direction, the critical thinking and the vigilance. I stand behind every claim and conclusion.)
