A Patch Tuesday Unlike the Others
On May twelfth, 2026, Microsoft released fixes for 118 vulnerabilities, sixteen of them rated critical. That volume alone would ordinarily read as an emergency. Instead, it was notable for what was absent: no actively exploited zero-days, and no previously disclosed flaws left unpatched. For the first time in nearly two years, the supply of fixes had gotten ahead of the supply of exploits.
The flaws themselves were serious. A stack-based buffer overflow in Windows Netlogon could hand an attacker SYSTEM privileges on a domain controller without any user interaction. A remote code execution bug sat in the Windows DNS client. Another flaw allowed an attacker to bypass Entra ID authentication by presenting forged credentials. None of these were exotic edge cases. They were the kind of foundational identity and access flaws that, left unpatched, quietly erode the trust systems every organization depends on.
The reason so many flaws surfaced at once has a name: Project Glasswing, an AI-driven vulnerability discovery tool developed by Anthropic and shared with a few dozen major vendors. Apple, Google, Mozilla, and Oracle all shipped unusually large patch batches around the same window, each crediting the same underlying capability.
What Changes When Machines Find the Flaws First
Glasswing functions like an inspector who can walk every inch of a building's frame at once, catching load-bearing defects a human reviewer would need months to find, if they found them at all. Apple's iOS update, which normally addresses roughly twenty vulnerabilities per release, fixed fifty-two, and Apple backported the changes as far back as the iPhone 6s. Firefox 150 resolved 271 vulnerabilities tied to the Glasswing evaluation and has since moved to a weekly security release cadence. Chrome's fix count jumped from thirty in April to one hundred twenty-seven in May.
This is, on its face, good news. Defenders are closing gaps faster than attackers can weaponize them, at least for this cycle. But the underlying story is not really about code. It is about what happens when the same class of technology that finds software flaws is also available to anyone looking to exploit the people who use that software.
The Human Side of the Same Machine Intelligence
In the same period that vendors were shipping record patch volumes, a separate analysis found that fraudsters using generative AI had driven a hundredfold increase in synthetic identities and a sevenfold rise in deepfake impersonations. These are not vulnerabilities in a codebase. They are vulnerabilities in the processes humans use to verify who they are talking to, who is applying for an account, and whose voice or face is on the other end of a call.
Synthetic identity fraud does not rely on stealing a real person's information outright. It assembles plausible-looking identities from fragments, real and fabricated, that pass automated verification checks because the checks were built for a world where creating a convincing fake identity took real effort. Deepfake impersonation attacks that same trust in a different place: the assumption that a familiar voice or face on a video call is reliable evidence of identity. Both techniques target the same underlying weakness, which is that human verification, and much of the automated verification built to mimic it, was never designed to withstand mass-produced, high-quality fakes.
This is the inversion worth sitting with. AI is currently helping vendors patch code faster than attackers can exploit it. At the same time, AI is helping attackers manufacture false identities and false faces faster than verification processes can catch them. The same underlying capability, applied to a different target, produces opposite outcomes for defenders.
Why the Patch Timeline Doesn't Tell the Whole Story
A zero-day-free Patch Tuesday is a genuine win, and it deserves to be recognized as one. But it addresses a narrower slice of the attack surface than it might appear to. Software vulnerabilities are a known, bounded category: they exist in code, they can be scanned for, and they can be fixed with a patch. Identity fraud and impersonation exploit something far less bounded: human judgment, trust, and the verification workflows built around them. No patch cycle closes that gap, because there is no single vendor responsible for shipping the fix.
That means organizations that measure their security posture primarily by patch compliance are only tracking half the picture. The other half is whether employees, help desks, and verification systems can still tell a real person from a synthetic or deepfaked one. As AI-assisted exploit development matures on the offensive side and AI-assisted discovery matures on the defensive side, the contest increasingly comes down to which side gets better use of the tools first, and identity verification is currently the side under the most pressure.
The Takeaway
The May 2026 patch cycle shows that AI can genuinely strengthen defense when it is pointed at code. The same month's fraud data shows that the identical technology, pointed at people, is scaling synthetic identities and deepfake impersonation faster than most verification processes can adapt. Patching software is necessary but no longer sufficient. The next layer of defense has to be built around how humans and the systems they rely on confirm identity, because that is where the machine intelligence gold rush is currently paying off for attackers.
If this kind of analysis is useful to you, subscribe free at threatlevelhuman.substack.com, and grab the free Social Engineering Red Flags field guide while you're there.
<- back to the blog