Microsoft Credits Its Own AI Tools for Rising Vulnerability Counts

TL;DR

Microsoft’s September 2026 Patch Tuesday — its monthly scheduled batch of security fixes — closed somewhere between 966 and 974 of its own security flaws, depending on which security-research firm counted them, plus dozens more from other vendors patched the same day. Two of the fixed flaws were already being used by attackers before Microsoft shipped the repair. Microsoft’s security team says its own AI-based scanning tools are finding more bugs, faster, than manual code review ever did, and expects that to keep pushing monthly patch counts higher. Independent researchers agree the volume is rising, but disagree on what that means: one warns more patches don’t automatically mean more real-world danger, another says it’s mostly an old backlog of hard-to-find bugs finally getting cleared.

What happened

How many flaws Microsoft actually fixed this month depends on who is counting. BleepingComputer put the number at 966, including 105 rated critical. Qualys counted 974 Microsoft vulnerabilities, 113 of them critical and 860 rated important, and separately noted 170 Adobe vulnerabilities, 50 of them critical, patched the same day. CrowdStrike counted 972 total flaws addressed, with 113 critical. The Hacker News, citing analysis from Trend Micro’s vulnerability-research program, the Zero Day Initiative, put the Microsoft total at 974 plus 25 from other vendors — 999 flaws in total.

Two of this month’s fixes were for flaws attackers were already exploiting before Microsoft shipped a repair — what the industry calls zero-days. One, tracked as CVE-2026-81963, let an attacker who already had limited access to a machine grant themselves full control through a flaw in the Windows Update mechanism itself. The other, CVE-2026-85880, was a memory-corruption bug in a Windows component that programs use to talk to each other on the same machine, which also let an attacker escalate their access.

Technically, Qualys broke the release down by how a flaw could be abused: 253 flaws (82 critical) would let an attacker run their own code on someone else’s machine remotely, 438 (27 critical) let an already-present attacker grab higher-level access, 173 (2 critical) could leak information, 56 could crash or disrupt a system, 16 (1 critical) let an attacker impersonate something trusted, and 19 (1 critical) could bypass an existing security check.

Microsoft’s own explanation centers on tools it built, not on outside attackers. Tom Gallagher, Vice President of Engineering at the Microsoft Security Response Center, wrote in April 2026 that AI-based scanning “AI can discover more issues, more quickly, across a broader surface area than previous methods.” He added that “Because these systems can work 24x7, limited only by available resources, we will discover a greater volume and diversity of vulnerabilities.” One example: Taesoo Kim, Microsoft’s Vice President of Agentic Security, described a system code-named MDASH in May 2026 that orchestrated more than 100 specialized AI agents to propose candidate bugs and then check each other’s work before reporting them, finding 16 vulnerabilities — four of them critical, remote-code-execution flaws — in Windows networking and authentication code.

Separately, Jeremy Chapman, a Director on the Microsoft 365 team, said in September 2026 that “If you’re not delivering critical quality updates with security fixes until a couple of weeks after they’ve been issued, that’s ample time for attackers using AI to find and exploit known security gaps.”

What this means (and what it does not)

Microsoft’s public framing turns a large patch count into a story about its own security investment: better AI-assisted scanning is catching more bugs before anyone else can. Microsoft’s Security Response Center itself makes this argument, and it is a framing that serves Microsoft — a rising patch count could just as easily read as a sign of accumulated flaws in existing code that better scanning is now surfacing. The security firms quoted alongside this release — Tenable and Fortra among them, plus Qualys and CrowdStrike — sell vulnerability-prioritization products, and urgency around record-sounding patch counts is also good for their business.

Those same researchers push back on treating a bigger number as automatically more dangerous. Satnam Narang of Tenable said “AI-assisted vulnerability discovery in 2026 is creating larger haystacks, but it isn’t finding more needles.” arguing the count of flaws that can actually affect most organizations stays low even as the raw total climbs. Tyler Reguly of Fortra offered a different explanation entirely: “Eventually, all those long-standing, hard to find vulnerabilities will be fixed and Patch Tuesday will return to its typical cadence.” framing this as a temporary backlog clearing rather than a new normal.

What does not follow from any of this is that September’s high count reflects a measured rise in AI-powered attacks against Microsoft or its customers. The AI explanation Microsoft itself gives is about its own internal scanning tools finding old bugs faster — a separate claim from any documented spike in attacker activity, and one that neither Gallagher’s nor Chapman’s statements establish.

What we still do not know

The exact size of this release cannot be pinned to one figure. Independent trackers disagree by up to 8 on the Microsoft-only total — 966 to 974 — and by up to 8 on the critical count, apparently because they use different rules for what counts as a distinct flaw and whether Extended Security Updates or non-Microsoft advisories are included. No source in this reporting could independently confirm a single authoritative total.

We also don’t know what share of this month’s fixes came from AI-assisted tools like MDASH rather than traditional manual research. Microsoft’s only disclosed MDASH results to date are the 16 vulnerabilities from one specific scan — a small fraction of any of the totals above.

No source resolves, with data, whether Narang or Reguly has the more accurate read on what rising patch counts mean for real-world risk. And no source in this reporting addresses how organizations are supposed to test and roll out roughly a thousand patches in a single monthly cycle, or what risk that pace of deployment itself creates.