Demo creative for Header leaderboard (728x90)

How the surge in AI-discovered bugs is straining Microsoft security and threatening tech stability

GeekWire
Image: GeekWire

Photo by Clint Patterson on Unsplash

It has come to the attention of the general public this month that adversaries successfully adopt technology a lot faster than defenders do. This has nothing to do with the willingness of defenders to try. It has everything to do with the ancient truth that to succeed, defenders have to get it right every time but adversaries only have to get it right occasionally.

Now multiply “every time” and “occasionally” by AI. Do you like that math? Does anyone?

The disastrous update that broke Teams, Outlook, SharePoint and other packages was not a direct attack by AI-wielding bad guys. Nor was the FAA outage . Nor were most of the outage events of the past few weeks, though who can say what’s up with those European Union refrigerators .

No. The true AI-driven mayhem has been predicated for several months now not on powerful new AI-honed attacks, but on an absolute deluge of bugs of relatively low severity that Microsoft and other defenders with a formal patch process aren’t built to process.

The question is whether traditional patch programs can adapt their processes to handle the deluge. Microsoft’s is MAPP (Microsoft Active Protections Program), to which the company has been committed for nearly two decades. If the industry-spanning events of the past few weeks are any indication, these programs can’t.

Microsoft CEO Satya Nadella said Wednesday at the New York Times DealBook Summit that last week’s Microsoft 365 outage came from a latent bug and wasn’t directly tied to AI, but he warned that AI-driven patching could bring more instability.

He elaborated in a separate appearance on CNBC the same day, saying the flood of patches being applied to older infrastructure is making systems less stable.

AI has “allowed us to find these bugs and fix these bugs, but that means you’ve got to then patch them, and that creates instability,” Nadella said, when asked by Jim Cramer about a string of recent outages at United, Schwab, Chase and other organizations.

Inside Microsoft’s patch machine

Some context: MAPP turned 18 this year. The program, run by the MSRC (Microsoft Security Response Center), handles outreach to partner companies that gain early access to information on the bugs in advance of each month’s Patch Tuesday release. In-house, it coordinates the patch release process, just as the larger MSRC organization helps coordinate the handling of bug reports from external finders.

MSRC also runs point on prioritizing patches of the worst bugs through its SSIRP (Software and Services Incident Response Plan) program, itself dating to 2004. In its time, it was an industry leader in how companies can efficiently get their arms around bad bugs.

MAPP in its own way is similarly remarkable. It has undergone a few traumas over the years, usually when a partner breaks confidence and leaks the advance details it receives. In 2025, after attacks on SharePoint raised suspicions of a leak from the program, Microsoft stopped giving Chinese companies early access . But MAPP continues to tick along in a way that would make any program manager proud. Other companies have their own version of MAPP / MSRC, as do community efforts such as those supporting Linux.

Full disclosure: I’m a former MSRC employee who learned about MAPP at the knee of the program’s principals. I subsequently tracked patch releases at a company partnered with MAPP at its highest (Validate) level, which means I was privy to the 14-day multi-stage run-up to each month’s release. I admire and respect the program… and I’m afraid that in the AI-finder era, it’s in big, big trouble.

A flood of bugs since April

Let’s look at the numbers for Microsoft.

Sixty months of Patch Tuesday release volume, from December 2021 to November 2026. OOB (out-of-band) releases are not represented unless the OOB involved a re-release related to a CVE included in the regular patch cadence. Click for larger image.

Since January 2021, Microsoft has issued patches for an estimated 10,319 CVEs. (I use estimated numbers throughout this analysis because there are various ways of counting such things, including whether or not the patch accrues to an application such as Edge, which actually relies on the Chromium project for its fixes. Other analysts arrive at larger or smaller counts.)

Over half of those — 5,383 CVEs — are 2026 releases, 5,128 of which were addressed since April. In comparison, 2025’s full-year CVE total was 1,180.

What happened in April? As nearly as analysis can determine, the floodgates opened and finders (security researchers) started barraging the company with reports.

The first quarter of 2026 saw 255 CVEs addressed over three months. April had 163, May 132; in June the number jumped to 209. In July it was off to the races with 575 CVEs. In October, the monthly count topped 1,400.

Most of these bugs aren’t coming from the usual security stalwarts. They’re coming from new finders, they’re coming in with multiple simultaneous finders not believed to be working together, and they’re coming in relatively weak.

Though the overall volume has skyrocketed, the number per month of bugs under active exploit in the wild — the notorious “zero-days” — hasn’t increased proportionally.

Moreover, average bug severity has in fact decreased. Until 2026, the average CVSS severity score for all bugs tended to hover around 7.8 (out of 10) each month. In 2026, that’s dropped to approximately 7.4 as of September — not garbage finds, but the sort of thing that doesn’t scramble the jets.

No surge in zero-days and a drop in severity — that’s good news. The problem is the volume.

A hand-crafted process

Once bugs are received by Microsoft, each goes through a MAPP-helmed, multi-team process of investigation and patch development.

@import url(‘https://fonts.googleapis.com/css2?family=Roboto+Slab:wght@500&display=swap’);@media (max-width:768px){.gw-pq{display:none!important}}

The true AI-driven mayhem has been predicated not on powerful new AI-honed attacks, but on a deluge of bugs of relatively low severity that Microsoft and others aren’t built to process.

Since many products have several versions still in support, that often means that multiple teams within each product group are asked to ingest a bug report, replicate the problem, devise and test a fix, ensure the fix doesn’t trigger other problems or inadvertently point to a similar bug elsewhere in the product and ready it for mass consumption. Even AI slop has to be taken seriously, just in case.

Meanwhile, MAPP is the traffic cop for the entire release, working with each team to ascertain patches are ready, understand the nature of the bugs and what can be safely communicated to partners and the public, and ready the processes that make the second Tuesday of each month a festival of downloads and reboots. From where we stand right now, it’s all fairly… hand-crafted.

In theory the process is automatable to some extent, though the success of that effort so far in light of recent events is debatable. But high volume also leads to other problems, such as exposing information on incomplete patching.

For instance, Microsoft has recently patched Office bugs that apply to both Windows and Mac, but patches have initially only been available for Windows versions.

Microsoft’s release-day information about such bugs — patching Windows versions while asking Apple users to monitor Microsoft communications for news of their own patch — alerts the world, including eager adversaries, that the flaw exists. Apple customers just have to hope no adversaries exploit it while their patch is pending.

More broadly, many of the steps in the current patching process involve judgment calls, not processes that scale well or that can reliably be handled by existing AI-driven systems. In Microsoft’s case, MAPP is a remarkably small team of humans and, being security folk at Microsoft, they’re not always the loudest or most popular voice in Redmond.

Add the mercurial nature of finders, who may or may not provide full details on their bugs (ask any MSRC person how often they receive a crash dump and a demand for payment) and may or may not disclose their finds publicly if they don’t feel they’re getting sufficient response.

Do not add the kind of defender staffing needed to handle a massive increase in volume — either in MAPP or in the individual product groups building the actual patches — because in the AI era, it’s likely the companies won’t add it either.

Emergency fixes pile up

The cracks are showing. Microsoft has been compelled to issue OOB (out-of-band) releases and servicing updates monthly since July, often within just a few days of the original release.

OOBs are important enough to merit release outside the usual patch cadence, often because there’s a security issue involved, but they’re tough on administrators already wading through a slew of regularly scheduled patching. With volumes high, patch testers may rush or elide certain steps, such as checking that a fix doesn’t break something deeper in the system.

Nadella told the DealBook audience this week that a patch for a critical bank system might once have been tested for a week. Now, he said, the patch “needs to go instantly from being available to being patched in production because it’s being exploited in the wild. So that’s going to create instability.”

Think of MAPP as Mariners catcher Dan Wilson in 2001: a pro handling a staff of skilled human pitchers, one pitch at a time. Now it’s 2026, and the pitchers have been swapped for a bank of mediocre pitching machines all slinging fastballs at … 2026-era Dan Wilson.

Here, again, AI can’t fix a system that requires defenders to catch every pitch. Secondary lines of defense that have worked in the past — for instance, the requirement by various regulatory programs that enterprises ascertain that patches are safe for their own systems — rely heavily on companies sending out the soundest possible patches.

Managed service providers might at least make sure that patches are flagged to customers as available, relevant and urgent. But they’re in no position to do much more than trust that what’s shipped is safe to deploy and react when things go wrong.

Even that gets harder when a bug kneecaps the security tools they depend on. Last week’s Microsoft 365 outage affected parts of Microsoft Defender that interact with Exchange Online.

It’s not just Microsoft, of course. The Linux community is seriously suffering , and both Apple and (especially) Google have struggled with patch-release quality in 2026 as evidenced by their own emergency updates and post-release device instability reports. Industry-wide, ecosystems built around reliable, transparent patch cadences and patch-development processes have never seemed more in danger of being overwhelmed.

AI alone won’t fix it

What needs to happen? Sure, build those AI systems; the marketing people will be pleased. It’ll speed up some parts of the process. But a system capable of replicating, every time, the human judgment calls that go into the patch-release process isn’t going to magically create itself; people have to create it.

Is there some chance that the deluge will ebb — that researchers and their tools will at some point finally find all the bugs in all the world’s products?

It’s a pretty thought, and some relief may eventually come. When more and more unrelated finders turn up the same bug, it gives one hope that the set is finite. But no one knows when relief will come, and most folk in tech suspect that really, no, there is no bottom to the bug barrel. Nor, for the foreseeable future, any rest for the weary.

Demo creative for Between results (336x280)