Attackers now exploit software flaws before most organizations know the flaws exist. Mandiant's M-Trends 2026 puts the median time from vulnerability disclosure to first exploitation at minus 7 days, so the typical exploit lands a week before the public advisory does. In 2018 that figure was 771 days. No patching program, however disciplined, can close a gap that opens before the race starts. The same AI that collapsed the window is also giving ordinary software teams the kind of security review that only a few elite companies could afford until recently. The organizations that come out ahead will stop trying to outrun attackers and start making the race irrelevant to their survival.
How fast are attackers actually moving in 2026?
Attackers now move faster than any human review cycle can follow. Microsoft's Digital Defense Report 2026 puts the median time from a flaw being found in the wild to weaponization at well under 24 hours. VulnCheck found that 32.1% of known-exploited vulnerabilities in the first half of 2025 were exploited on or before their CVE disclosure day, up from 23.6% a year earlier.
The driver is cost.
The cost of a working exploit has fallen to pocket change — so attackers no longer choose targets.
A University of Illinois study showed that GPT-4, running as an autonomous agent, exploited 87% of tested one-day vulnerabilities when given nothing but the CVE description. AI frameworks now reproduce 51% of all published CVEs with working exploits at an average cost of about $2.77. A working proof-of-concept takes 10-15 minutes and roughly $1. One agent swarm found more than 100 exploitable kernel flaws in 30 days for about $600 total. When an exploit costs less than a cup of coffee, attackers stop choosing targets carefully and simply try everything they can reach.
The exploit window has collapsed below zero — the median exploit now lands 7 days before the public advisory. Source: Mandiant M-Trends 2026.
Volume compounds the problem. Microsoft counts about 40,000 CVEs published in the first half of 2026, with the year tracking toward roughly 72,000. In July 2026 came JADEPUFFER, the first fully automated ransomware attack. Its entry point was ordinary: an internet-exposed service running known-vulnerable software. No zero-day, no nation-state tradecraft, just a machine that found an open door faster than anyone could close it.
Why can't we just patch faster?
Defenders are getting slower while attackers get faster, and the gap is now measured in weeks on one side and minutes on the other. The Verizon DBIR 2026 reports that the median full-patch time rose to 43 days from 32, and only 26% of critical known-exploited vulnerabilities were fully remediated, down from 38%. Qualys measured the mean time to remediate complex enterprise applications at 5 months and 10 days, with about 45% of vulnerabilities still unpatched after a year.
None of this reflects lazy teams. Patching an insurer's claims platform means regression testing against every broker integration that touches it. A mine's processing plant may run control software that can only be updated during a scheduled shutdown. A logistics operator cannot take its warehouse management system offline in peak season. A portfolio company acquired last quarter may be running servers nobody on the current team installed. These constraints are real, and none of them shrink because exploits got cheaper.
| Measure | Attacker velocity | Defender velocity |
|---|---|---|
| Time to act | Working exploit in 10-15 minutes | Median full patch in 43 days |
| Starting point | Exploitation median at minus 7 days from disclosure | Response begins at disclosure |
| Coverage | 51% of published CVEs reproduced with working exploits | 26% of critical known-exploited flaws fully remediated |
| Cost and scale | 100+ kernel flaws found in 30 days for ~$600 | ~45% of vulnerabilities unpatched after a year |
| Trend | Weaponization well under 24 hours | Patch time up from 32 days to 43 |
Minutes versus weeks: attacker weaponization against the median enterprise patch cycle.
Read across any row and the conclusion is arithmetic. Shaving days off your patch cycle still reduces risk, and you should do it. It will never bring the exposure window to zero, because for a growing share of flaws the window opens before a patch exists.
Is AI only making this worse for defenders?
AI is also building what we call the defense dividend: software that ships with security review most companies could never have paid for before. The capability that writes a $1 exploit can read a codebase and flag the same flaw before release. Microsoft expects vendors using AI to patch their own code to ship measurably more secure software once the initial wave passes.
Until recently, continuous expert review of every code change belonged to the largest software houses, because it required specialists who were scarce and expensive. Today a mid-sized vendor building fleet telematics or a tokenization platform can run AI review against every pull request. The flaws an attacker's agent would find get found by the vendor's agent first, and fixed before customers ever install the code.
Adoption is already broad. The WEF Global Cybersecurity Outlook 2026 found that 77% of organizations use AI in security operations, with phishing detection at 52% and intrusion and anomaly response at 46%. Executives also feel the pressure: 87% ranked AI-related vulnerabilities the fastest-growing cyber risk, and 94% expect AI to be the most significant driver of change.
The defense dividend: the same AI that writes $1 exploits reviews every pull request before release.
The dividend arrives unevenly, though. The near term will be rough, because AI is surfacing a backlog of old flaws faster than vendors can repair them. After that, new code should get harder to break. Your job is to survive the backlog period and position your organization to collect the dividend once it pays out.
If patching can't keep up, what actually protects us?
Reducing what attackers can reach and stopping exploitation at runtime is what protects you now. Patching remains essential hygiene, but your primary defense shifts to two questions. Can an attacker reach this system at all? If they can, will the system detect and block the exploit's behavior even while the underlying flaw is unpatched?
Runtime protection works because an exploit for an unknown flaw still has to do something once it lands. It has to spawn a process the application never spawns, or open a connection to a server the attacker controls. Behavioral monitoring, network segmentation and virtual patching at the network edge all target that behavior rather than the specific flaw. A web application firewall rule can block the malicious request pattern for a vulnerability your team has not patched yet. Segmentation means a compromised claims server cannot reach the payment ledger. Endpoint detection can kill a process that suddenly starts encrypting files or dumping credentials, whatever flaw let it in. None of these controls require you to know the CVE number in advance, and that is the point. You stop betting your survival on winning a race you have already lost, and start betting on containing the damage when the exploit arrives first.
What should leaders do first?
Start with exposure, then evidence, then speed, then people. The order matters because each step makes the next one possible.
- Inventory every internet-facing asset this month. You cannot protect what you do not know you own. JADEPUFFER got in through an exposed service running known-vulnerable software, and that pattern repeats across most automated attacks. Run external attack surface scans, reconcile the results against your asset register, and pay particular attention to recently acquired subsidiaries, forgotten test environments and vendor-managed systems. Anything exposed that does not need to be exposed should come off the internet before you do anything else.
- Prioritize by exploitation evidence, not severity scores. A CVSS 9.8 flaw nobody is exploiting matters less than a 7.5 flaw on the CISA Known Exploited Vulnerabilities catalog. With roughly 72,000 CVEs expected this year, ranking by severity alone guarantees your team spends weeks on theoretical risks while active ones sit open. Use KEV listings, exploit prediction scores and threat intelligence feeds to decide what gets fixed first, and accept that some lower-risk items will wait.
- Set a 5-7 day patch SLA for exposed systems and back it with runtime protection. For internet-facing systems with known exploitation, the 43-day median is indefensible. Commit to five to seven days, measure it, and report it to the board. Where a system genuinely cannot be patched in that window, such as plant control software waiting on a scheduled shutdown, require a documented compensating control: virtual patching, isolation, or behavioral monitoring tuned to that asset. The SLA makes speed accountable. The runtime layer covers the gap the SLA can never close.
- Invest in the people who run all of this. The WEF Global Cybersecurity Outlook 2026 found that 54% of organizations name a lack of knowledge and skills as the main barrier to using AI in security. Buying AI tools without people who can tune, question and act on their output just produces a faster stream of ignored alerts. Fund training for your existing analysts, give engineers time to learn AI-assisted code review, and treat security staffing as operational capacity rather than overhead.
Frequently asked questions
Should we stop prioritizing patching? No. Patching still removes risk permanently, and it remains your baseline hygiene. What changes is that you can no longer rely on it as your primary defense for exposed systems, because a growing share of exploits arrive before patches exist.
What is the CISA KEV catalog and why does it matter more than CVSS? The Known Exploited Vulnerabilities catalog lists flaws with confirmed exploitation in the wild. CVSS measures how bad a flaw could be in theory. When you have tens of thousands of CVEs and limited staff, confirmed attacker activity is the better signal for what to fix first.
Is runtime protection only for large enterprises? No. Managed endpoint detection, cloud web application firewalls and basic network segmentation are available to mid-sized organizations at reasonable cost. A logistics operator or regional insurer can deploy meaningful runtime controls without building a security operations center from scratch.
When will the AI defense dividend actually pay off? Expect the backlog period to last while AI keeps surfacing old flaws in legacy code, likely measured in years rather than months for older platforms. New software from vendors using AI review should improve sooner. Ask your key suppliers now whether they run AI-assisted security review on every release, and weight their answers in renewal decisions.
Where this leaves you
My view is that the industry spent too long treating patch speed as the headline metric, and the data from 2026 has made that framing obsolete. A minus 7 day exploitation median is not a problem better process will solve. The organizations I would bet on are the ones that shrink their exposed footprint aggressively, assume some exploits will land, and design their environments so a single compromised service cannot become a company-wide outage. That is a less satisfying story than "we patch faster than anyone," but it matches what attackers are actually doing.
The longer-term picture is more hopeful than the current numbers suggest. The same economics that make a $1 exploit possible make continuous code review affordable for software teams that never had it, and that should steadily raise the cost of attacks over the next several years.
None of this happens automatically. It happens because someone in the organization decides to pull the asset inventory, challenges the severity-only prioritization, pushes for a tighter SLA on exposed systems, and argues for the training budget. If you are reading this with the authority or the influence to start any of those four steps, you are that person, and the work you start this quarter will determine which side of the next automated attack your organization lands on.


