PATCH PULSE / Cadence Observatory 2011–2026
Microsoft CVEs / yr Zero-days exploited in the wild ×8 All CVEs published ÷ 40
MIXED SCALES · 2026 = YTD/FORECAST · tap any point for its components & sources

The headline signal

Two hypotheses are commonly offered for the growing patch burden: vendors are shipping more often, or each release is getting bigger. The data supports both — but unevenly. Cadence has accelerated (five major vendors formally sped up between 2023 and mid-2026), while the payload per release has grown far more. 2026 is the inflection year, and every major analyst attributes the surge to AI-assisted vulnerability discovery.

Vendor channels tracked
39 · IT + OT
Five categories, from hyperscalers to grid-control vendors. AI VERIFIED
Largest single drop ever
622 CVEs
Microsoft, July 2026 — triple the June record; counted 622 / 621 / 569 by different trackers. AI VERIFIED
Microsoft · annual CVEs
909 → 1,130
2023 → 2025 (+24%); 2026 passed every prior full year by July. AI VERIFIED
All CVEs published / yr
18k → 48,185
2018 → 2025; ~66,000 forecast for 2026 (FIRST). REPORTED
Zero-days exploited ITW
90 in 2025
Within the 60–100/yr baseline that has held since 2021 (pre-2021 norm: 20–30). AI VERIFIED
CISA ICS advisories 2025
508 · record
Carrying 2,155 CVEs; average severity CVSS 8.07. AI VERIFIED
Cadence accelerations
5 vendors
Chrome (2023), Oracle, Adobe, Cisco, Mozilla (2026) all formally sped up. None slowed down. AI VERIFIED
Loudest new channel
Linux kernel
~290 → 3,529 → 5,679 CVEs/yr after becoming its own CNA (Feb 2024). AI VERIFIED
AI vuln-research entities
56 indexed
Programmes, labs, platforms and threat actors — the engine behind the surge (chapter 06). REPORTED
Why counting is now contested

The same July 2026 Microsoft release was counted as 622 (Microsoft's Security Update Guide), 621 (ZDI) and 569–570 (Tenable, excluding CVEs resolved through servicing or owned by other vendors) — with ~480 further Chromium/Edge bugs alongside. The Linux kernel, meanwhile, assigns a CVE to nearly every bug fix on principle. Any cross-vendor comparison needs a declared counting basis; this tool flags the basis and confidence on every figure. And the divergence is now structural: from July 2026, Microsoft's own Security Update Guide stopped itemising individual CVEs — publishing a summary table by product family plus a "Notable CVEs" shortlist — so even the primary source must now be reconstructed by third-party trackers.

Microsoft: annual CVE volume AI VERIFIEDPRE-2017 ESTIMATED

Tenable/ZDI Patch-Tuesday counting from 2017. Faded bars are estimates; 2011–16 bars are security bulletins, a different unit shown only for era context. 2026 is year-to-date at 16 July — already the tallest bar.

The ecosystem behind it: all CVEs published REPORTED

All CVE identifiers published per year (NVD/CVE Program; pre-2024 approximate). The dashed final bar is FIRST's 2026 forecast (~66,000). NVD submissions rose 263% between 2020 and 2025, forcing NIST to re-prioritise enrichment in April 2026.

Vendor directory

Thirty-nine channels across five categories — including the operational technology vendors whose controllers, safety systems and SCADA platforms run physical infrastructure. Filter by category, search by name, and select a card to open the full signal: cadence, release size, zero-day pressure, key events and sources.

Compare vendors

Side-by-side telemetry. Click any column header to sort. "Drops/yr" is the scheduled cadence in mid-2026; "Volume" is each vendor's own counting unit — the units differ, which is why the Method notes matter.

The great acceleration — scheduled drops per year, 2024 → mid-2026 AI VERIFIED

Dots show 2024 cadence; arrowheads show mid-2026. Adobe cited "AI-accelerated vulnerability discovery compressing the disclosure-to-exploitation window" when doubling to twice-monthly on 14 July 2026; Cisco moved to risk-based twice-monthly disclosure the same month. No tracked vendor has reduced cadence in fifteen years.

Comparison table

Vendor Category Drops / yr (2026) Volume (own unit) Zero-day pressure Trend

Volume figures use each vendor's own counting basis and are not directly comparable — see Sources → Method.

Zero-day pressure

"Zero-day" here follows the GTIG/Mandiant definition: a vulnerability maliciously exploited in the wild before a patch was publicly available — the vendor was fixing it after attackers were already using it.

Industry-wide zero-days exploited in the wild AI VERIFIED

2021 shows Google's combined TAG+Mandiant figure (106); Mandiant alone counted 80 that year. 2026 is GTIG's mid-year pace estimate (~88).

Who carried the exploited zero-days AI VERIFIED

GTIG attribution by vendor. Microsoft tops the list every year tracked. 2024 total 75 (later revised to 78); 2025 total 90.

2025 target split: the enterprise pivot AI VERIFIED

Enterprise technologies took a record 48% of 2025 zero-days (43 of 90); security and networking appliances alone accounted for roughly half of those. Browser zero-days fell to eight — the hardened targets are working, so attackers moved to the edge.

Zero-day fixes per year — three well-tracked channels MIXED BASES

Apple and Chrome are journalist-maintained running counts of exploited-in-the-wild fixes; Microsoft is Tenable's Patch-Tuesday zero-day count (disclosed or exploited pre-patch — 41 in 2025, of which 24 were exploited). Apple's 2023 spike (20) was the Pegasus / BlastPass / Operation Triangulation year. Microsoft 2021–23 not charted: no comparable series located.

Fixed after release — the n-day problem

Zero-days are only half the "fixed after the horse bolted" picture. The window from patch release to exploitation of the patched flaw had shortened to about five days by 2023 (Ivanti threat data), and patch-diffing is now automated — attackers diff a new build against the old one to locate the fix and weaponise it before most estates finish testing. In 2025 CISA added 244 entries to its Known Exploited Vulnerabilities catalogue (total 1,483): the closest public proxy for "shipped, then exploited anyway". Notable n-day disasters: ESXiArgs ransomware (2023) exploited a two-year-old VMware patch gap; Citrix Bleed kept paying ransomware crews months after its fix.

Cadence map

Each channel below is one calendar year of release rhythm (January → December). Tick height suggests typical drop size. The stories live in the rhythm changes — and between 2023 and 2026 the whole board sped up.

Release rhythm — representative year per channel

Cadence change-log AI VERIFIED

OCT 2003

Microsoft creates Patch Tuesday — second Tuesday monthly; the model the industry copied. Adobe, SAP and Ivanti all later aligned to the same day.

2005 →

Oracle settles on quarterly CPUs — third Tuesday of Jan / Apr / Jul / Oct, a rhythm held for two decades. Juniper adopts the same quarterly pattern for scheduled JSAs.

2011 →

CISA's ICS advisory programme scales — 67 advisories (103 CVEs) in its first full year becomes 508 advisories (2,155 CVEs) by 2025, with average severity climbing from CVSS 6.44 to 8.07. The OT giants — Siemens, Schneider Electric, and later Rockwell, AVEVA, Phoenix Contact, ABB, Mitsubishi and Moxa — progressively converge on an "ICS Patch Tuesday" aligned with Microsoft's day.

2011

Firefox moves to rapid release — from annual versions to a ~6-week (later 4-week) train, normalising fast browser shipping.

AUG 2015

Android goes monthly — the Stagefright disclosures push Google onto committed monthly bulletins; Samsung mirrors with its monthly Security Maintenance Release.

31 DEC 2020

Flash Player retired — the decade's most-patched product leaves the board; browser exploit-kit activity collapses with it.

JAN 2023

CISA stops mirroring Siemens advisories — the volume outgrows the shared feed, and the sector's authoritative source begins fragmenting: by 2025 only 22% of published OT/ICS vulnerabilities carry a CISA advisory (58% in 2024).

AUG 2023

Chrome adds weekly security updates (from Chrome 116) on top of ~4-weekly milestones, explicitly to shrink the n-day exploitation gap.

FEB 2024

Linux kernel becomes a CVE Numbering Authority — and starts assigning a CVE to nearly every fix: ~290 CVEs in 2023 becomes 3,529 in 2024 and the largest CNA volume of any organisation by 2025.

MAY 2026

Oracle adds Critical Security Patch Updates in the 8 non-CPU months — from four drops a year to effectively twelve.

Average CVEs per Microsoft monthly drop AI VERIFIED

2024 avg ~84/month; 2025 ~94 (records: Jan 157, Oct 167); 2026 H1 averaged over 200 with June ~206 and July 622. The release day hasn't changed — the payload has.

Linux kernel CVEs per year — the CNA effect AI VERIFIED

Same codebase, new counting policy: after kernel.org became its own numbering authority (Feb 2024) and adopted an "almost every fix gets a CVE" stance, volume rose ~12×. A caution for anyone benchmarking vendors on raw CVE counts.

AI Signal — the engine behind the 2026 surge

Every 2026 anomaly in this observatory — Microsoft's 622-CVE month, five cadence accelerations, a forecast 66,000 CVEs — traces to the same cause: AI-assisted vulnerability discovery went operational. This tab maps who is doing the discovering, incorporating the companion research index Zero-Day AI-Assisted Observatory (June 2026 edition): 56 programmes, labs, platforms and threat actors, each carrying its own hedged claims and source links. Tap any bar to list the entities behind it.

Capability curve
<9% → ~50%
UK AISI-reported success rate of frontier models on apprentice-level cyber tasks, 2023 → 2025. REPORTED
Autonomous find-and-fix
86% / 68%
DARPA AIxCC: share of synthetic bugs found, and of those patched (~45 min average). All seven systems open-sourced. AI VERIFIED
The mirror image
May 2026
GTIG reports the first confirmed AI-assisted zero-day exploited in the wild, linking agentic tooling to DPRK- and PRC-linked actors. REPORTED

Who is hunting — 56 entities by category COMPANION INDEX

Tap a bar to open every entity in that category, with claims and source links. Vendor research teams dominate by count, but the three frontier-lab programmes account for the largest reported finding volumes.

When they arrived — launches by year COMPANION INDEX

The long tail of foundations (CERT/CC 1988, OSS-Fuzz 2016, bug-bounty platforms) — then the 2023–26 cluster. Seven entities launched in the first half of 2026 alone. Tap a bar to see that year's arrivals.

The find-and-fix frontier

Anthropic — Project Glasswing (Apr 2026): reports 10,000+ high/critical flaws found with ~150 vetted partner organisations; claims the public Claude Security patched 2,100+ flaws in three weeks.
OpenAI — Daybreak (May 2026): reports "Squidbleed" (CVE-2026-47729), a 29-year-old Squid flaw; runs Patch the Planet across 19 open-source projects with Trail of Bits.
Google — Big Sleep & CodeMender: Big Sleep reportedly caught an in-the-wild SQLite flaw (CVE-2025-6965) pre-exploitation — an AI first; CodeMender is said to have upstreamed 72 open-source fixes in six months.
Microsoft — MDASH: the agentic scanner already feeding Patch Tuesday, finding 16 of May 2026's bugs alone.
AISLE / XBOW: AI-native start-ups — AISLE claims all 12 zero-days in one OpenSSL release; XBOW reportedly topped the US HackerOne leaderboard autonomously.

The strain — and the mirror

The maintainer bottleneck: curl shut its HackerOne bounty in January 2026 citing "AI slop" (reportedly under 5% of reports valid) — then genuine reports roughly doubled, and the seven-person team took July 2026 off entirely. ~60% of open-source maintainers are unpaid.
The doctrine: the NCSC's patch-wave guidance argues AI in skilled hands exploits accumulated technical debt at scale, warning of a "forced correction" (NCSC's argument, as catalogued in the companion index).
The threat side: GTIG reports the first confirmed AI-assisted zero-day used in the wild (May 2026), with agentic tooling linked to state-nexus actors. "Every defensive programme on this page has a mirror image" is editorial framing of that finding.

Full index — 56 entities

Incorporated from the companion research page Zero-Day AI-Assisted Observatory (June 2026 edition). Entries reproduce that index's own hedged language — "reports", "claims", "said to" — and its source links; treat vendor-reported figures accordingly.

Insights — the questions this tool set out to answer

Direct answers, with the hedges honest data demands. Provenance: statements below are the compiler's aggregation of figures cited elsewhere in this tool (every underlying number is linked in the relevant tab, in the tap-to-inspect panels, and under Sources). Direct quotations and single-source claims carry inline links; sentences without a link are analytical synthesis of that linked data. Where a judgement goes beyond the sources, it is flagged as interpretation.

1 · Which vendors ship the most patches and significant fixes? AGGREGATED — SOURCES INLINE

By unique CVEs fixed in its own products, Microsoft dominates: 1,130 in 2025, more than 1,250 by mid-July 2026 — already a record year. By raw patch count, Oracle is the heavyweight (~1,379 patches across 2025's four CPUs — a sum of the quarterly figures linked in its vendor card), though ~75–80% cover open-source components rebundled inside Oracle products (October 2025: 374 patches ↔ 170 unique CVEs). By identifier volume, the Linux kernel now tops everyone — thousands of CVEs a year since 2024 — but that reflects a counting philosophy, not a security collapse (interpretation, consistent with the kernel CVE team's own description of its policy). By ecosystem blast radius, Google spans Chrome (weekly security fixes cascading into Edge, Opera, Brave and every Electron app), Android (monthly, frequently 40–120 CVEs) and cloud. These leader-boards measure different things; treat any single ranking with suspicion.

In plain EnglishMicrosoft fixes the most bugs in its own products. Oracle ships the most patches — but mostly for open-source code it bundles. The Linux kernel prints the most CVE numbers because it counts differently. There is no single “biggest patcher”: the answer changes with what you count.
2 · What is each vendor's cadence? AGGREGATED — SOURCES INLINE

Monthly, second Tuesday: Microsoft (since 2003), SAP, Ivanti, and Adobe until July 2026 (now twice-monthly). Monthly bulletins: Android and Samsung (since Aug 2015), Atlassian. Quarterly: Juniper and F5 (Quarterly Security Notifications); Oracle until May 2026 (now effectively monthly). Weekly security refreshes: Chrome (since Aug 2023), Mozilla (since spring 2026). Continuous severity-driven streams: Red Hat errata, IBM bulletins, the Linux kernel's stable releases. Rolling/event-driven: Apple, VMware, Citrix, Fortinet, Palo Alto Networks, SonicWall, Check Point, Progress — and most OT vendors, where Siemens, Schneider Electric, AVEVA and Phoenix Contact hold a monthly ICS Patch Tuesday while GE Vernova, Emerson, Yokogawa, Omron, Hitachi Energy and Delta Electronics publish rolling advisories largely through the CISA channel. Cisco ran semi-annual bundles plus rolling advisories until July 2026's twice-monthly model. The direction of travel is one-way: five formal accelerations since 2023, and no formal deceleration among tracked vendors in fifteen years (aggregated observation across the cadence change-log; sources per entry there).

In plain EnglishAlmost every major vendor now ships fixes on a published schedule — and the schedules only ever get faster. In fifteen years, nobody has slowed down.
3 · Have patches grown in size — more vulnerabilities per release? AGGREGATED — SOURCES INLINE

Yes, decisively — this is the stronger effect. Microsoft's monthly average went ~75 (2023) → ~84 (2024) → ~94 (2025) → 200+ (2026 H1 — derived from the monthly figures in the Microsoft card), with July 2026's 622 tripling any pre-2026 month. Oracle's quarterly CPUs run 300–380 patches (per-quarter sources in its card). SAP's Patch Day notes rose ~39% year-on-year into 2025. The backdrop: ~18,000 CVEs published in 2018 vs 48,185 in 2025 (+165% — computed from those two figures), with ~66,000 forecast for 2026. Cadence roughly doubled in places; payload per release grew far more.

In plain EnglishThe big change isn't how often patches arrive — it's how much is packed inside each one. Microsoft's monthly bundle has grown roughly two-and-a-half times in two years, and July's held 622 fixes.
4 · How many zero-days, and how often? AGGREGATED — SOURCES INLINE

Industry-wide exploitation stepped up in 2021 and has stayed elevated since: from a 20–30/year norm (2018: 16; 2019: 32; 2020: 30 — Mandiant) to a 60–100/year baseline (2021: 106 combined; 2022: 63; 2023: 100; 2024: 78 (75 at publication); 2025: 90 — GTIG). At the 2025 rate, a new in-the-wild zero-day surfaces roughly every four days somewhere in the ecosystem (arithmetic on the GTIG total, not a GTIG statement). Per vendor in 2025: Microsoft 25, Google 11, Apple 8, Cisco 4, Fortinet 4. Microsoft's own Patch Tuesday data shows the strain — 22 zero-days in 2024 → 41 in 2025 (24 exploited) — the "roughly two per month" framing is this tool's arithmetic, not Tenable's. Apple's worst year was 2023 (20, overwhelmingly commercial-spyware chains); Chrome runs steadily at ~8–10; the network-edge vendors (Ivanti, Cisco, Fortinet, Citrix, Palo Alto, VMware) supplied the defining zero-day incidents of 2023–2025.

In plain EnglishA decade ago attackers used maybe two dozen unpatched flaws a year. Now it's about ninety — one new one roughly every four days — and increasingly aimed at VPNs and firewalls, the very boxes bought for protection, rather than browsers or phones.
5 · What else does the data show? AGGREGATED — SOURCES INLINE

a) AI is now the named driver. ZDI ("AI is supercharging flaw discovery at an uncontrollable scale"), Microsoft (expect sustained higher volumes; its MDASH agentic scanner found 16 of May 2026's bugs alone), Adobe (cited AI-compressed windows when doubling cadence) and the UK NCSC (April 2026 warning) all converge. "The discovery curve bent with the LLM build-out" is this tool's synthesis of those statements — the full ecosystem doing the discovering is mapped in the AI Signal tab.

b) The target moved from consumer to enterprise edge. Enterprise products took a record 48% of 2025 zero-days, with security and networking appliances around half of those, while browser zero-days fell to eight (GTIG). "The hardened consumer targets are working; the edge is not" is interpretation consistent with GTIG's own analysis.

c) Triage infrastructure is under visible strain. NVD submissions rose 263% (2020→2025); NIST formally re-prioritised enrichment in April 2026 and moved older backlog to "Not Scheduled". The shift toward EPSS scores plus the CISA KEV catalogue rather than CVSS-first triage is widely advocated in the linked commentary; the conclusion that severity labels alone no longer discriminate reflects the severity distributions in the linked Microsoft and OT datasets, and is interpretation.

d) Patch speed cuts both ways. Automated patch-diffing means every larger drop is also a larger exploit menu (established practitioner knowledge; see the linked commentary). The release-to-exploit window was ~5 days by 2023, per Ivanti threat data. "Deployment speed, not just patch availability, is now the variable that matters" is this tool's conclusion from those inputs.

e) Counting is contested. 622 vs 621 vs 569 for one release; a kernel that CVEs nearly every bug fix by policy; an Oracle patch count dominated by rebundled open source. The recommendation — that any metric built on "vulnerabilities patched" needs a declared counting basis — is this tool's, drawn from those documented divergences.

In plain EnglishAI now finds software flaws faster than humans can process them; the systems that count and rank those flaws are straining; and every bigger patch is also a bigger treasure map, because attackers compare new code with old to see what changed.
6 · What does the picture look like in operational technology? AGGREGATED — SOURCES INLINE

Cadence has converged on IT's rhythm; deployment hasn't. Siemens, Schneider Electric, AVEVA, Phoenix Contact and increasingly Rockwell, ABB, Mitsubishi and Moxa now publish on an "ICS Patch Tuesday" aligned to Microsoft's day, while the DCS and grid houses — Honeywell, Emerson, Yokogawa, GE Vernova, Hitachi Energy — remain rolling and service-programme driven. But an advisory is not a deployed patch: controllers and safety systems patch during planned outage windows measured in months, a meaningful share of advisories ship mitigations rather than fixes for long-lived product lines, and open-source component fixes can take years to reach industrial firmware.

Volume and severity are both climbing. CISA's ICS advisories grew from 67 (103 CVEs) in 2011 to a record 508 (2,155 CVEs) in 2025, with average severity rising from CVSS 6.44 to 8.07 and 82% of 2025 advisories rated high or critical. Critical manufacturing (~2,000 cumulative advisories) and energy (1,400+) are the most affected sectors, and Purdue Level 1 devices — PLCs, RTUs, IEDs — top the asset list.

The shared feed is fragmenting. Only 22% of 2025's published OT/ICS vulnerabilities carried a CISA advisory (58% in 2024, 40% in 2023), 134 vendors published flaws with no ICSA at all, and CISA stopped mirroring Siemens advisories in January 2023. "Anyone tracking OT risk through a single feed is seeing a shrinking fraction of it" is the direct implication of those figures — interpretation, but a short step.

And the defining OT attacks barely used CVEs. Stuxnet needed IT zero-days to reach the PLCs, not to program them; Triton abused legitimate Triconex engineering functions; PIPEDREAM weaponised native protocols; FrostyGoop (2024) spoke plain Modbus to heating controllers in Lviv; the 2023 Unitronics water-utility defacements used default credentials on internet-exposed PLCs (PIPEDREAM: CISA advisory AA22-103A; other incidents per the linked vendor cards). The closing judgement — that patch metrics measure one dimension of OT risk, with exposure, segmentation and insecure-by-design protocols carrying the rest — is this tool's interpretation, aligned with the cited incident record.

In plain EnglishFactory and grid vendors now announce fixes as fast as the software industry — but installing them still waits for planned shutdowns. And the worst industrial attacks didn't need unpatched bugs at all: they used legitimate features and default passwords.

Fighting back — the defence stack

The story so far is pressure: bigger drops, faster exploitation, an AI-driven discovery surge. This chapter is the response — the defences that demonstrably work, traditional and innovative, arranged as five rings from prevention to resilience. Tap any ring or card for the evidence and sources. As throughout, aggregation is labelled and vendor-reported claims keep their hedges. AGGREGATED — SOURCES INLINE

The defence stack — five rings, outside in AI VERIFIED CORE

Read from the outside in: prevent flaws existing, prioritise the ones that matter, patch faster than exploitation, contain what gets through, and engineer so that even success cannot become catastrophe. Tap a ring for its evidence.

What the data says works AI VERIFIED

The strongest prevention evidence anywhere in this tool: the share of Android vulnerabilities caused by memory-safety bugs, before and after Google moved new code to Rust — 76% → 24% → below 20%, against an ~70% industry norm.

Four more working defences, measured:
· Browser hardening: exploited browser zero-days fell from ~15 (2021) to 8 (2025) — sandboxing, site isolation and fast update rings did that.
· Attack-surface removal: Flash retired (Dec 2020) → browser exploit-kit activity collapsed. The cheapest patch is the component you removed.
· No-reboot patching: Windows Server 2025 hotpatch — up to 8 no-reboot months a year, free via Arc since May 2026 — removes the main reason patches wait.
· Machine-speed fixing: DARPA AIxCC systems patched 68% of found flaws in ~45 minutes, all open-sourced.

The traditional arsenal — proven, still load-bearing

Tap any card for evidence, caveats and sources.

The innovative edge — where the curve gets bent

Maturity flags: PROVEN measured at scale · SCALING deployed and growing · EMERGING promising, evidence still forming.

Where the standards put these defences INDICATIVE MAPPING

Every ring of the defence stack already lives inside the major frameworks — NIST CSF 2.0, the UK's Cyber Assessment Framework, IEC 62443 for industrial systems, the CIS Controls and ISO 27001. Tap any cell for the specific control references, any column header for the framework, or a ring name for its evidence. Mappings are a navigation aid compiled by the AI research compiler, not a certified crosswalk.

Read the pattern, not just the cells: no ring maps to a single clause, and no framework is satisfied by a single ring. The frameworks were written to demand a system. And sector overlays sharpen them further: GB energy applies the CAF through the published DGE Sector CAF Overlay, which tunes these outcomes for OT reality — sector-specific expectations on vulnerability management and compensating controls sit on top of everything in this matrix.

⚠ The point-response trap — basics first, system always

Everything in this chapter is a point response to a specific pressure — and point responses do not sum to security. Three disciplines hold them together. First, the basics, everywhere: the unglamorous floor codified as Cyber Essentials and CIS Implementation Group 1 — asset inventory, access control, timely updates, backups. The incident record in this tool (default credentials at water utilities, two-year-old unpatched hypervisors, internet-exposed controllers) is overwhelmingly a record of missing basics, not missing innovation. Second, think in outcomes, not tools: the CAF and CSF 2.0's Govern function are deliberately holistic — they ask whether risk is actually managed across people, process, suppliers and technology, precisely so that buying a defence cannot substitute for achieving one. Third, layer for failure: every ring in the stack assumes the ring outside it will sometimes fail. Memory-safe code will not stop a phished credential; a hotpatch will not save an unsegmented network; and nothing on this page removes the need to rehearse recovery. That layered, honest posture — not any single innovation — is the actual defence. And fourth, scan the horizon while you do the basics: the next attack vectors are already forming inside AI-built and AI-operated systems themselves — prompt injection and tool abuse against agentic AI, poisoned components in AI-assisted supply chains, and autonomous agents wired into workflows their owners don't fully understand. Companion research by the same author maps these surfaces in depth (Agentic AI Attack Surfaces; The Always-On Agent Land Grab) — the basics protect today's estate; horizon scanning protects the estate you're about to build.

In plain EnglishThe defences that measurably work share one idea: stop racing the flood and change the river. Write code that can't have whole classes of bugs. Patch running systems without the reboot excuse. Let machines fix what machines find. And for grids and factories — design the physical process so a hacked computer still can't cause harm. But none of it works as a shopping list: the standards all ask for the same thing — basics done everywhere, layers built as a system, and a plan for the day a layer fails.

Sources & method

Every figure traces to a link below. Confidence flags: AI VERIFIED — figure checked by the AI research compiler directly against a primary source or named analyst team (Tenable RSO, ZDI, GTIG, Oracle, NIST, kernel.org); REPORTED consistent across ≥2 reputable secondary sources; ESTIMATED pre-2017 or methodology-fragile figures — directional only. Cross-referencing: Microsoft counts checked Tenable ↔ ZDI ↔ Microsoft SUG; zero-days checked GTIG ↔ Mandiant ↔ vendor advisories; ecosystem counts checked NVD/NIST ↔ CVE Program ↔ CVE.ICU.

Publication note — independence, speculation & corrections

This tool is an independent research compilation with no vendor sponsorship, and nothing in it constitutes an endorsement or criticism of any vendor — larger disclosure counts frequently indicate stronger transparency. Speculation policy: the tool aims to publish verified or clearly attributed figures only; where a statement is analytical synthesis it is labelled AGGREGATED with the reasoning shown, where a count is a detection floor it says floor, and where a figure is an estimate it carries the red ESTIMATED flag and should not be quoted as fact. Claims sourced from vendor-published research (including AI-lab programme results) are reported in the original publisher's hedged language and have not been independently validated here. Corrections: figures were verified against the linked sources at compilation (16 July 2026); sources may revise their numbers, and corrections to this tool are welcomed and will be applied with a change note.

Change note — revision 12.1, 18 July 2026: post-publication verification pass. GTIG's 2023 zero-day figure standardised to 100 (97 at original publication); 2024 stated as 78. July 2026 detail expanded: third zero-day (BitLocker CVE-2026-50661, publicly disclosed), diverging critical counts (59–62), Security Update Guide format change, and the August-embargoed SharePoint chain — with Rapid7, CrowdStrike and Malwarebytes added as sources. UI responsiveness pass across small, large and touch/keyboard environments.

Method & known limitations

1. Counting bases differ. Microsoft figures use Tenable RSO's Patch-Tuesday CVE counts (ZDI differs by ±1%); Oracle counts patches, not unique CVEs, unless stated; SAP counts security notes (one note may cover several CVEs); the Linux kernel CVEs nearly every fix. Cross-vendor rankings are therefore indicative only.
2. Pre-2017 Microsoft data is not CVE-comparable — bulletins were the unit (est. 135 in 2015, 155 in 2016) until the Security Update Guide transition.
3. Zero-day counts are floors, not totals — GTIG notes it can only count detected and disclosed exploitation. Apple/Chrome per-year tallies are journalist-maintained running counts.
4. 2026 figures are year-to-date (to 16 July) or labelled forecasts.
5. Edge-vendor series are qualitative — Cisco, Fortinet, Palo Alto, Citrix and Juniper publish rolling advisories without clean annual CVE series; their cards lean on documented incidents and GTIG attribution. A future version could derive per-vendor annual series from the CISA KEV catalogue and NVD CPE data.
6. OT vendor cards measure disclosure, not deployment. In operational technology an advisory routinely precedes fleet-wide patching by months or years (outage windows, certification, safety cases), a visible share of advisories are mitigation-only, and the defining OT attacks abused insecure-by-design protocols and credentials rather than fresh CVEs. Zero-day charts for OT vendors are floors based on public reporting.
7. Order-of-magnitude figures. Red Hat and IBM advisory volumes are stated as orders of magnitude ("thousands"/"hundreds" per year) because exact counts depend on product-set definitions; per-advisory data is available from the linked primary errata/bulletin hubs. For several OT vendors (GE Vernova, Emerson, Yokogawa, Omron, Delta Electronics), advisories publish primarily through CISA's ICS channel, so the CISA hub is cited as the working primary source.
8. AI Signal entries are incorporated from a companion research index (Zero-Day AI-Assisted Observatory, June 2026) and reproduce its hedged language and links verbatim; they have not been independently re-verified here.
\n 9. This is an independent research compilation, not a vendor scorecard; larger CVE counts often indicate more transparent reporting, not weaker engineering.

AI built from public sources - no trackers or analytics - 18 July 2026