Microsoft's September security release was enormous. Ars Technica counted roughly 972 patched vulnerabilities, including 112 rated critical. Two months earlier, the record was around 570.
The tempting conclusion is that Microsoft software suddenly got much worse, but that's not case. A more useful interpretation is that the industry's ability to find bugs is changing at the same time attackers are gaining faster tools to exploit them.
To be clear, Microsoft has not said all 972 September vulnerabilities were discovered by AI, but it's reasonably to speculate that many, if not most, were. The record patch count and Microsoft's AI-security program are separate facts. Together, however, they show the same operational pressure: find and fix more problems before attackers can turn disclosed flaws into working attacks.
Microsoft already has AI agents hunting bugs
Microsoft's codename MDASH system orchestrates more than 100 specialized AI agents across an ensemble of models. Those agents perform different jobs inside a structured pipeline: inspect code, surface suspicious paths, validate exploitability, deduplicate findings, and help move real bugs toward fixes.
In May, Microsoft said MDASH helped its researchers find 16 vulnerabilities in Windows networking and authentication components, including four critical remote-code-execution flaws. The company later expanded the system into Defender and GitHub security workflows, and Microsoft Learn now describes a production pipeline that scans, validates, deduplicates, and can help generate code fixes.
The important object here is not one brilliant model. It is the harness around the models.
A security agent needs more than raw reasoning. It needs access to the right code, specialized context, validators that can prove a finding is real, a way to eliminate duplicates, and a workflow that lands with a human owner who can ship a patch.
That architecture is why Microsoft can swap in newer or cheaper models without rebuilding the whole security operation.
The patch gap is becoming a race
On The Neuron's September 9 livestream, Corey Noles described a security concept worth remembering: every public patch starts a second clock.
Defenders get a fix. Attackers also get a clue.
A motivated attacker can compare the patched version with the old one, work backward to understand the vulnerability, then search for systems that have not updated yet. Historically, parts of that reverse-engineering process could take substantial manual effort.
AI can compress pieces of the loop. Models can read diffs, inspect large codebases, propose exploit paths, write test code, and iterate faster than a human doing every step by hand.
Security lab Calif offered a vivid demonstration this week with WeWorm, a research project that created a zero-click worm capable of spreading through WeChat calls on iOS and Android. The researchers said AI helped them find the memory-corruption flaw in two days and build the full worm in under a week. They reported the issue to Tencent before publishing the demo.
That is the uncomfortable symmetry of AI security: the same automation that helps defenders find vulnerabilities can also help attackers understand and weaponize them.
What security teams should change
The response is not to panic every time Patch Tuesday gets bigger. It is to shorten the distance between disclosure and action.
First, know what you own. Fast patching is impossible without a live inventory of endpoints, servers, packages, and internet-facing software.
Second, rank exposure instead of treating every CVE equally. Internet reachability, sensitive data, privileges, and whether an exploit is already circulating matter more than the raw number of patches.
Third, make dependency checks part of normal development, especially when coding agents can add packages faster than humans can review them.
Grant Harvey described a simple version of that habit on the livestream: before installing a third-party library, he asks an AI system to search for recently documented supply-chain attacks or compromises involving that package. That is an extra sanity check, not a replacement for package signing, vulnerability scanners, or a real security review.
Time-to-fix becomes the metric
A higher vulnerability count can actually be good news if it means defenders are discovering and closing real holes earlier.
The number to watch is how quickly a validated bug moves from discovery to a patched production system, and whether attackers can consistently move through the same loop faster.
Microsoft's own AI-security work points in that direction. In May, MDASH found concrete Windows bugs. By August, Microsoft had published the scanner as part of a broader agentic code-security workflow. September's record patch volume does not prove AI found the month's 972 vulnerabilities, but it shows the scale of the job security teams are being asked to do.
AI is giving both sides faster tools. The advantage will go to whichever side builds the shorter, more reliable path from signal to action.
