Moving faster is more expensive

On September 8, 2026, Microsoft shipped the largest Patch Tuesday in its history: 974 CVEs, more than any single release the company has ever issued, blowing past the previous record of 570 CVEs set just two months earlier in July. Microsoft has openly attributed part of the surge to MDASH, a multi-model agentic scanning harness it now uses internally to hunt for vulnerabilities, and it has warned customers this is the new normal, not a one-off spike.

Two weeks earlier, WordPress’s plugin ecosystem hit its own version of the same wall. Patchstack’s 2026 security whitepaper describes a “significant increase” in AI-generated vulnerability report “slop”, reports that are incomplete, invalid, or simply wrong, submitted in volume because AI makes it trivially cheap to generate something that looks like a finding. Patchstack now runs a dedicated triage team just to separate real vulnerabilities from noise before they ever reach a plugin vendor.

Different ecosystems, same pattern: AI has made finding things dramatically cheaper. The question this piece asks is what that has done to the other half of the ledger, the cost and speed of actually fixing what’s found.

The discovery side: the numbers are not subtle

• The NVD had logged 60,475 vulnerabilities by early September 2026, more than all of 2025 (48,185), with four months of the year still to go.
• FIRST’s mid-2026 forecast projects roughly 66,000 CVEs for the full year, citing AI-assisted discovery, a 449% year-over-year jump in GitHub Security Advisory volume, and a 3,119% increase in “CNA of last resort” activity absorbing a backlog of previously unassigned vulnerabilities.
• Microsoft’s own Patch Tuesday trajectory in 2026: 164 (April) → 200 (June) → 570 (July) → ~400 (August) → 974 (September).

This is discovery accelerating because the cost of looking has collapsed. That was the expected story going in. The less-told story is what happened downstream.

The remediation side: it’s not keeping up. It may be getting slower

• Verizon’s 2026 DBIR found the median time to fully remediate a critical vulnerability rose to 43 days, up from 32 the year before. Only 26% of CISA KEV entries were fully remediated in 2025, down from 38%.
• Qualys’ April 2026 enterprise benchmark put mean time to remediation for complex applications at 5 months and 10 days, while, over the same stretch, attacker weaponization speed has compressed roughly 6x since 2022, from 32 days down to 5.
• Mandiant’s M-Trends 2026 report describes mean time-to-exploit as having gone negative: roughly -7 days, meaning exploitation now routinely precedes patch availability, versus a 63-day cushion defenders had in 2018.
• Contrast Security’s AppSec Overflow 2026 report: the average production application carries 22 critical vulnerabilities; teams close 3.4 per month. At that rate a backlog doesn’t shrink, it compounds. A critical flaw located in a live environment takes 92 days on average to remediate.

Put those two halves side by side and you get the actual finding: discovery cost has fallen off a cliff, and remediation cost, measured in calendar time, which is the proxy for headcount and process cost, has risen, not merely failed to keep pace. That’s the opposite of what a naive “AI helps security” narrative would predict, and it’s a stronger, more falsifiable claim than “AI increases risk.”

Patches aren’t landing clean, either

Your instinct that “reliable patches take longer” shows up in the data too, from two directions:

AI-generated patches fail more than half the time. 1Password research cited in recent coverage found AI-generated vulnerability patches failed to fully fix complex software flaws 53.9% of the time, meaning a meaningful share of “resolved” tickets are actually latent reopens waiting to happen.

Real-world WordPress cases of exactly that pattern:

• Burst Statistics (200,000+ active installs): an authentication-bypass flaw, CVE-2026-8181, was introduced in version 3.4.0 and was still present in the very next release, 3.4.1. It wasn’t actually closed until 3.4.2, two release cycles after the vendor believed they’d fixed it, and after Wordfence had already logged over 7,400 blocked attacks in 24 hours.
• church-admin plugin: a researcher publicly documented bypassing the patch for an arbitrary file upload vulnerability twice, the fix shipped, was defeated, was fixed again, was defeated again.
• Abandoned Cart Lite for WooCommerce (CVE-2023-2986): version 5.15.0 was believed to fix an authentication-bypass flaw; it didn’t. The actual fix landed in 5.15.2.
• WordPress core itself: the “wp2shell” pre-auth RCE chain disclosed in July 2026 combined two bugs, one of which (CVE-2026-63030) was a regression – a bug reintroduced in version 6.9 in code that had existed safely since 2020. It was mass-exploited within a day of disclosure.

There’s no single industry-wide statistic tracking “what percentage of patches get bypassed or reopened”, it isn’t a metric anyone publishes systematically, which is itself worth a line in the piece. But the pattern in individual cases is consistent: a patch is not the end of the timeline, it’s a new checkpoint that itself needs verification, and every one of those checkpoints costs the vendor and every downstream site owner another remediation cycle.

The counterargument worth including, for balance

Not every AI-found vulnerability matters, and treating raw CVE volume as a 1:1 proxy for risk overstates the story:

• VulnCheck’s H1 2026 report found that of 1,061 vulnerabilities attributed to AI-assisted discovery, only 14 (1.3%) have been confirmed exploited in the wild, roughly matching the baseline exploitation rate for all vulnerabilities, AI-found or not.
• Of the more than 23,000 findings Anthropic’s Project Glasswing surfaced in open-source code, only 126 resulted in published CVEs, and just one has been confirmed exploited.
• FIRST frames this as “Rain vs. Flood”: total CVE volume is way up, but the subset that’s actually exploited or credibly exploitable hasn’t grown at anywhere near the same rate.

This actually sharpens the article’s thesis rather than undercutting it. The cost isn’t rising because more of what’s found is dangerous, it’s rising because triage now has to do more work to find the small signal in a much larger pile of noise, and every one of those triage hours is a remediation-side cost that didn’t exist three years ago.

What this adds up to

1. Discovery got radically cheaper. Attack surface, Microsoft’s product line, WordPress’s plugin ecosystem, open source broadly, is being scanned at a volume and speed that didn’t exist two years ago.
2. Remediation got slower and more expensive, not faster, on every metric that tracks time-to-fix: Verizon’s 43 vs. 32 days, Qualys’ 5-month-plus enterprise MTTR, Contrast’s 92-day critical-flaw average, all moving the wrong direction.
3. The gap between weaponization speed and patch speed inverted. Mandiant’s -7-day mean time-to-exploit means the old model, disclose, patch, then worry about exploitation, no longer describes most incidents.
4. A nontrivial share of “fixed” is actually “believed fixed.” Whether it’s a 54% AI-patch failure rate or a WordPress plugin needing three release cycles to actually close a hole, the remediation clock frequently restarts.
5. Volume itself is a cost. Sorting signal from AI-generated noise, Patchstack’s “slop” reports, or Microsoft’s own 974-CVE release needing prioritization before any of it gets touched, is a new labor line that scales with discovery, not with actual risk.

The honest framing isn’t “AI made us less safe”, VulnCheck’s numbers argue against that. It’s that AI moved the bottleneck. It used to sit at discovery. It now sits at triage and remediation, and nobody has resourced that side of the pipeline to match.

---

Sources

• Infosecurity Magazine / The Register / SecurityWeek / SecurityAffairs — Microsoft September 2026 Patch Tuesday coverage (974 CVEs)
• Tenable — Microsoft July 2026 Patch Tuesday (570 CVEs, MDASH)
• Patchstack — State of WordPress Security in 2026 whitepaper
• VulnCheck — State of Exploitation 1H-2026
• Verizon — 2026 Data Breach Investigations Report
• Qualys — April 2026 enterprise patch benchmark (via Cloud Security Alliance whitepaper)
• Mandiant — M-Trends 2026
• Contrast Security — AppSec Overflow 2026
• FIRST — 2026 Mid-Year Vulnerability Forecast
• Cloudflare — September 2026 NVD CVE count citation
• BleepingComputer — Burst Statistics (CVE-2026-8181) coverage
• Medium / Patchstack bug bounty write-up — church-admin plugin patch bypass
• GitHub — CVE-2023-2986 (Abandoned Cart Lite for WooCommerce)
• harizanov.com — WordPress core “wp2shell” regression writeup (CVE-2026-63030)
• 1Password research on AI-generated patch failure rates (via Contrast Security coverage)