Microsoft Skips a Zero-Day for the First Time in Two Years. Nobody Wants to Talk About Why.

Key points

• Microsoft patched 118 vulnerabilities this Patch Tuesday, none under active exploitation and none previously disclosed.
• Sixteen are rated critical, led by a Netlogon buffer overflow that grants SYSTEM on a domain controller with no credentials required.
• Apple, Mozilla, Oracle and Google all shipped unusually large patch volumes this cycle.
• Project Glasswing, an Anthropic code-auditing capability, is the thread connecting the surge. No vendor will say how many findings came from it.
• Anthropic has confirmed the project exists but won't release a participant list, a discovery count or methodology.

What broke the zero-day streak?

For the first time in roughly two years, Microsoft's May Patch Tuesday arrived without a single fix for a vulnerability already under active exploitation. All 118 patches also cover flaws that weren't previously disclosed, which matters because prior disclosure hands attackers a map. Microsoft's Security Response Center hasn't said what changed. Two follow-up emails went unanswered at publication.

Rapid7 flagged the three critical flaws worth prioritising. CVE-2026-41089 is a stack-based buffer overflow in Windows Netlogon: no credentials, no user interaction, low complexity, SYSTEM on a domain controller. Patches cover Windows Server back to 2012. CVE-2026-41103 is an Entra ID bypass via forged credentials, rated by Microsoft as "exploitation more likely." CVE-2026-41096, a DNS client remote-code-execution flaw, gets a "less likely" label from Microsoft, which used the same language for Exchange bugs in 2023 that were under attack within a fortnight. We first covered CVE-2026-41089 on 28 May, a day after Anthropic's own vulnerability-discovery numbers became public.

Should you worry about Project Glasswing?

Project Glasswing is an Anthropic-built code-auditing capability. A few dozen large software makers, Microsoft among them, have been given access to it. Microsoft acknowledged its participation in a single sentence and wouldn't say how many of this month's 118 fixes originated from Glasswing versus internal research or bounty submissions.

The volumes elsewhere are hard to ignore. Apple typically fixes around 20 vulnerabilities per iOS update cycle, according to Chris Goettl, vice president of product management at Ivanti. On 11 May, Apple shipped fixes for at least 52, backported to iOS 15 on the iPhone 6s. Apple's release notes don't credit Anthropic or any AI-assisted discovery. Mozilla resolved 271 vulnerabilities in Firefox 150, reportedly during the Glasswing evaluation, and has since moved to a weekly security cadence. Oracle's most recent quarterly update closed 450 flaws, more than 300 of them remotely exploitable without authentication, and the company has now switched to monthly critical updates. Google's Chrome update on 8 May patched 127 bugs, up from 30 the prior month.

None of these vendors will say, on the record, what share of those findings an AI surfaced. Our earlier reporting found that Anthropic's Mythos AI identified 23,000 potential vulnerabilities across 1,000 open-source projects, with confirmed critical findings and no public word yet on whether any were exploited before disclosure.

What Anthropic will and won't say

Anthropic's press team confirmed Glasswing exists. It declined to provide a participant list, a discovery count or a methodology document, pointing instead to a forthcoming research post with no date attached.

That's the part worth watching. If an LLM is finding hundreds of exploitable bugs per vendor per quarter in shipping code, the obvious question is who else is running the same playbook. Defenders are patching faster. The lag between a finding and a fix is shrinking. What nobody's measuring publicly yet is the lag between an adversary running a similar tool and that fix arriving.