Microsoft shipped fixes for 966 vulnerabilities in its September 2026 Patch Tuesday, the largest single release the company has ever put out and enough to carry its running total for the year to roughly 2,750 β more than double the previous annual record of about 1,250 set in 2020, according to InfoQ's tally. Two of the flaws were already being exploited.
Key takeaways
- BleepingComputer's breakdown counts 438 elevation-of-privilege bugs, 258 remote code execution flaws, 173 information disclosure issues and 105 rated Critical by Microsoft.
- The two exploited zero-days, CVE-2026-81963 in the Windows Update Stack and CVE-2026-85880 in Windows ALPC, are both local privilege escalation bugs carrying a CVSS score of 7.8.
- Microsoft has been running an AI-powered vulnerability discovery system internally, and security practitioners say the constraint has moved from finding bugs to evaluating, testing and deploying the fixes.
What actually shipped in September
Vendor tallies differ slightly because they draw the boundary around the release in different places β Tenable counted 964 CVEs and 104 Critical entries, while SecurityWeek put the broader figure at 974. The shape is consistent regardless of which line you take: privilege escalation dominates, remote code execution makes up roughly a quarter of the batch, and just over a hundred items carry Microsoft's Critical rating.
The two bugs under active attack are the ones with a clock on them. CVE-2026-81963 is a link-following flaw in the Windows Update Stack that lets an authorised attacker escalate to SYSTEM; it was reported by researcher Romain Deperne alongside the Microsoft Threat Intelligence Center. CVE-2026-85880 is a heap-based buffer overflow in Windows Advanced Local Procedure Call with the same outcome, credited to Volexity's Mark Kelly and to David Galazin and Jeremy Hedges at Proofpoint.
Neither is a remote entry point on its own. Both are the second half of an intrusion chain, which is exactly why they show up in real campaigns: an attacker who already has a foothold uses them to take the box.
Why the numbers keep climbing
Microsoft has confirmed it uses an AI-powered system to hunt for flaws across its own products, and the volume curve tracks that adoption. It is not an isolated case. Security journalist Brian Krebs has noted that Microsoft is far from the only large vendor shipping outsized patch bundles this year, with several companies crediting large language model tooling for the jump in findings.
Writing for Ars Technica, Dan Goodin placed the surge against an open letter in which OpenAI, Anthropic, AWS, Google, Microsoft and others warned that AI-enabled cyber attacks would become substantially more widespread and sophisticated. The industry response has been to point the same capability at its own code first β a defensible strategy that happens to generate a very large invoice for everybody downstream.
The pattern is already visible in offensive tooling as well, where autonomous testing agents have started surfacing findings that survived years of conventional review. Our earlier report on an AI pentester that located a three-year-old admin token in 25 minutes describes the same dynamic from the attacker's side of the table.
The complaint from the people who deploy it
None of the practitioners quoted in InfoQ's report disputed that finding more bugs is good. Their objection was about throughput at the other end of the pipeline.
Jack Bicer, director of vulnerability research at Action1, argued that the difficulty at this scale is not working through the list but knowing what deserves attention first, with IT teams forced to separate the items that demand same-day action from those that can ride the normal deployment cycle. Marva Bailer, founding CEO at Qualaix, put the operational cost more plainly:
Finding the problem is one step. Organizations still have to understand their exposure, test the patch, determine what else it might affect and then deploy it across potentially thousands of devices and interconnected systems.
Bailer also flagged the squeeze that matters most: AI shortens the gap between a flaw existing and a flaw being known, which compresses the window teams have to test and roll out a fix before someone weaponises it. Tyler Reguly, security R&D associate director at Fortra, was blunter still, suggesting that while Microsoft is playing catch-up on its own backlog the raw counts have stopped carrying much signal at all.
What changes for IT and security teams
The practical shift is that patch management is becoming a prioritisation problem rather than a coverage problem. Nobody is deploying 966 fixes in a maintenance window, so the work moves upstream: mapping which of those CVEs touch systems that are actually exposed, and building enough test automation to validate a batch without hand-checking each one.
That favours organisations with accurate asset inventories and automated rollout pipelines, and penalises everyone running patch cycles on spreadsheets. If the vendor curve holds β and Microsoft has given no indication it plans to slow discovery down β October will not be smaller.
FAQ
How many vulnerabilities did Microsoft patch in September 2026?
BleepingComputer counted 966 flaws fixed on Patch Tuesday, while Tenable logged 964 CVEs and SecurityWeek counted 974 across the wider release. The differences come from where each tracker draws the boundary around the release, not from a dispute about the underlying advisories.
Which September 2026 vulnerabilities are being actively exploited?
Two: CVE-2026-81963 in the Windows Update Stack and CVE-2026-85880 in Windows ALPC. Both are local elevation-of-privilege flaws rated CVSS 7.8 that let an attacker who already has access on a machine escalate to SYSTEM, so they should be prioritised ahead of higher-scoring but unexploited items.
Is AI responsible for the record patch volume?
Partly. Microsoft has confirmed it runs an AI-powered vulnerability discovery system, and other vendors have credited similar tooling for their own increases. Attribution for any individual CVE is not published, so the honest answer is that AI-assisted research is a major contributor rather than the sole cause.
