Computerworld’s report follows a ProPublica investigation into Microsoft’s work with Anthropic’s Claude Mythos Preview through Project Glasswing. ProPublica reviewed internal Microsoft material and a recording of a May meeting in which engineers described a “mad dash” to deal with findings generated by the model. In one SharePoint-focused presentation, Mythos reportedly surfaced 90 critical and 141 important bugs in April alone, with more arriving during the first half of May.
That is a serious volume problem. It is not, however, the same as a public disclosure that those bugs are exploitable on every Windows 11 PC, that they remain unfixed today, or that attackers have used them. ProPublica reported that Microsoft declined to say how many Mythos-identified bugs had been patched since the internal presentation, and it remains unclear whether any particular finding from the program has been exploited in the wild.
The evidence is broader than Windows — and narrower than the headline
The “Windows flaws” framing needs a correction. The clearest internal figures reviewed by ProPublica concern SharePoint, while the broader set of products named in the report includes Microsoft 365, Teams, and Copilot. Those are enormously consequential products, especially in enterprise and government environments, but they are not interchangeable with the Windows client operating system.
Microsoft has separately confirmed that AI-assisted security research is finding real bugs in Windows. In May, the company said its multi-model agentic scanning harness, called MDASH, helped researchers find 16 vulnerabilities across Windows networking and authentication components. Microsoft assigned CVEs to those findings, including critical remote-code-execution issues in the Windows TCP/IP stack and IKEv2 service; the National Vulnerability Database independently records at least one of those CVEs, CVE-2026-33827, as a Windows TCP/IP remote-code-execution vulnerability.
The important difference is timing and disposition. Microsoft’s public MDASH examples were vulnerabilities that had already moved through its security-release process and into a Patch Tuesday update. The internal Mythos material, by contrast, described an intake and remediation queue—especially around SharePoint and other services—where high-severity cases were arriving faster than teams could close them.
Microsoft’s July Windows security post acknowledges the larger trend without identifying a Windows-specific Mythos backlog. The company says AI will cause security releases to contain more fixes, because its systems can search more code, generate more candidate findings, validate more of them, and pass confirmed issues to engineers more quickly.
That is Microsoft’s stated strategy. The missing information is what matters most to customers: how many AI-discovered flaws are still open by product, what severity they have after human triage, whether they are reachable in default configurations, and whether any have indicators of active exploitation. Microsoft has not published that inventory.
Microsoft’s triage model is rational, but AI raises the cost of leaving the long tail open
According to ProPublica, Microsoft’s internal plan was to tackle critical findings first, then important ones, then hundreds of moderate-severity issues. That is conventional vulnerability management. A vendor cannot responsibly push a patch for every scanner result without confirming the flaw, determining its reachability, writing a fix, testing it against a huge number of supported configurations, and ensuring the cure does not introduce an outage or a new security defect.
The process is especially difficult for Windows and Microsoft 365. A patch may need to work across consumer PCs, managed enterprise estates, virtual desktops, servers, domain controllers, security products, legacy applications, drivers, and different servicing channels. Finding a defect can be automated more easily than proving exploitability and validating a fix across that matrix.
The concern raised by the ProPublica reporting is that the old severity ladder may be less reliable when advanced models can combine separately modest weaknesses into an attack chain. Vinh Nguyen, an Anthropic adviser and former NSA chief AI officer, told ProPublica that several low-level bugs can collectively produce high-impact access. Microsoft responded that chaining has long been considered in its vulnerability assessment and risk analysis, while also saying AI is forcing the industry to reconsider how it treats issues previously placed in lower priority buckets.
Both points can be true. Security teams have always chained bugs; the change is that a capable model may make the search for viable chains cheaper, faster, and more repeatable. A “moderate” flaw that needs unusual conditions may remain moderate on its own. It becomes a different operational risk if attackers can rapidly identify the three other conditions needed to turn it into an intrusion path.
That does not make a policy of patching every low-severity issue immediately practical. It does mean vendors need better ways to assess combinatorial risk: whether a flaw is likely to become dangerous when paired with other findings in the same product, deployment model, or identity environment.
Patch Tuesday’s record numbers do not prove Windows has become less secure
Microsoft’s July 14 security release was unusually large, but even the public count is unsettled. ProPublica described Microsoft as releasing fixes for more than 600 bugs. Windows Central, citing Action1’s analysis, put the figure at 570 vulnerabilities across Microsoft products. The discrepancy is a reminder that Patch Tuesday totals are not a precise health metric: different trackers may count CVEs, advisories, product variants, or bundled fixes differently.
Microsoft has not said that every July fix came from MDASH, Mythos, or another AI system. Windows Central explicitly noted that Microsoft did not claim MDASH found all 570 issues in that release. So it would be wrong to draw a straight line from a large monthly patch bundle to the internal backlog described by ProPublica.
Still, the direction is clear. More capable AI discovery increases the number of security issues that vendors can see, and that means customers should expect larger updates and more frequent urgent remediation decisions. A higher number of disclosed bugs can reflect better discovery and more active patching rather than a sudden collapse in code quality.
The uncomfortable part is that discovery helps attackers too. Anthropic’s testing, first reported by Axios, found that Mythos Preview could generate proof-of-concept exploits for disclosed Windows kernel vulnerabilities quickly, including a first proof of concept in 31 minutes in one test. Those tests involved already public vulnerabilities and patches, not secret Microsoft findings. But they reinforce the practical point: the time between public disclosure and usable exploitation is shrinking.
The traditional enterprise assumption that a patch can wait through a monthly test cycle is becoming harder to defend for internet-facing, high-value, or broadly deployed systems.
What Windows administrators should change now
For home Windows users, the action is simple: leave Windows Update enabled and install monthly security updates promptly, rather than treating a high CVE count as a reason to wait indefinitely. The reporting does not support panic-patching every optional preview release, but it strongly supports not allowing security updates to age for months.
For IT teams, the response should be operational rather than theatrical:
- Deploy Microsoft’s monthly security updates through phased rings, but make the first validation ring fast enough that critical fixes do not remain stalled in testing without a documented risk decision.
- Track patch compliance by exposed role and business impact, not just by the percentage of devices that eventually become compliant. Internet-exposed systems, privileged admin workstations, remote-access infrastructure, domain services, and systems holding sensitive data should have shorter remediation targets.
- Keep Microsoft Defender and other endpoint-security signatures current while updates move through testing. Microsoft says it uses Defender protections and the Microsoft Active Protections Program to reduce exposure between disclosure and full update deployment, but those layers are supplemental—not substitutes for applying the patch.
- Use management tools such as Windows Autopatch, Intune, and supported hotpatch capabilities where they fit the environment, while maintaining an update rollback and incident-response plan. Microsoft’s Known Issue Rollback can reverse certain problematic changes, but it cannot turn every bad update into a harmless event.
The main change is cultural: patching must move from a calendar exercise to a risk-based process with an explicit view of exposure. A system that cannot be updated quickly needs compensating controls—network isolation, stronger identity protections, application restrictions, or reduced administrative access—rather than a vague exception record.
The problem Microsoft has not quantified
Microsoft has been unusually candid in acknowledging that AI will increase vulnerability discovery and security-update volume. Its Windows team says it is adding AI to discovery, triage, fix development, regression analysis, and validation, while retaining human review. That is the only plausible long-term response: detection at machine speed must be matched by evidence gathering, engineering, testing, deployment, and mitigations that work at something closer to machine speed.
But the company has not released the number customers need to judge its progress: the age and severity distribution of confirmed AI-discovered vulnerabilities that remain unpatched across Windows, Microsoft 365, SharePoint, Teams, and Copilot. Nor has it tied individual public CVEs or Patch Tuesday releases back to Project Glasswing findings in a way that would show whether the backlog is shrinking.
So, should Windows users worry? Yes—about the patch gap, not an unproven Windows emergency. AI is making it easier to find and weaponize flaws, which makes current, well-managed Windows updates more important. The evidence so far shows Microsoft finding and fixing more bugs, while struggling to process the expanding queue; it does not show a disclosed, actively exploited cache of Mythos-discovered Windows vulnerabilities waiting on ordinary PCs.