Microsoft's Own Engineers Couldn't Keep Up With Anthropic's Bug-Hunting AI

In mid-May, a group of Microsoft engineers and managers logged onto a call, and into a conference room in Redmond, to talk about a problem most companies would love to have: an AI model was finding security bugs in their software so quickly that their own team couldn't fix them fast enough.

The effort is called Project Glasswing, and the AI model at the center of it is Mythos, built by Anthropic. According to a recording of the meeting viewed by ProPublica, Anthropic had given a small group of organizations, including Microsoft, early access to the tool so they could hunt for vulnerabilities in widely used software before criminals or hostile governments found the same holes first.

The stakes of that race are the whole point of the project. Software vulnerabilities are a currency in cyberwarfare and espionage. Nation-states including China have a long track record of finding and exploiting flaws in commercial software to spy on governments and companies. Anthropic's pitch, in effect, was that if an AI model could out-hunt human researchers, it made more sense to put that model in the hands of defenders first.

Based on what Microsoft's own engineers described in the meeting, that pitch checked out. One slide presented that day showed that in April alone, Mythos had surfaced 90 bugs in SharePoint, Microsoft's collaboration software used by companies and governments worldwide, that were rated "critical," plus another 141 rated "important." The pace didn't slow down. In just the first half of May, Mythos found even more.

That volume created a genuine scramble inside Microsoft. When one engineer on the call asked whether Mythos had "lived up to the hype that Anthropic claimed it would have had," a manager's answer was simple: "Yes." The tool was working almost too well. Engineering manager Hans Andersen told the group they were now in what a colleague described as "a mad dash" to close the gap between discovery and patching, urging the team to prioritize clearing out April's backlog of bugs: "Please, please, please if your org has any April bugs, drive those down."

A ticking clock

The urgency in that meeting came down to timing. Andersen explained that Microsoft's team believed they had roughly two weeks of exclusive access left before, as he put it, "the rest of the world will have caught up" — with May 31 marked as the cutoff.

The engineers on the call weren't shy about pressing on what that actually meant in practice. One of them distilled it into a blunt hypothetical: if similar bug-hunting capability became broadly available on June 1, would adversaries have access to Microsoft's own unpatched vulnerabilities by June 2? The exchange captured the core tension of the whole project. Anthropic's tool was a head start, not a permanent advantage, and once equivalent AI-driven vulnerability hunting became available more widely, the same techniques uncovering flaws for defenders could just as easily be turned by attackers.

Why this matters beyond one product

SharePoint isn't a niche piece of software. It sits behind document collaboration and file sharing at companies and government agencies around the globe, which is exactly why a wave of newly discovered "critical" and "important" bugs in it is significant. Every one of those flaws that Mythos found before an attacker did is a potential intrusion, data breach, or espionage campaign that never happens. Every one still sitting unpatched when outside access to similar tools catches up is a live risk.

The episode also says something bigger about where AI is headed in cybersecurity. For years, the promise of AI-assisted vulnerability research has been mostly theoretical, a capability vendors talked about more than one that reliably outpaced skilled human teams. What played out inside Microsoft's Project Glasswing meeting suggests that gap has started to close, and quickly: an AI system generating critical bug counts fast enough to overwhelm one of the world's largest software companies' own patching capacity is a notable marker, whatever comes next for Mythos itself.

It also underscores the double-edged nature of the technology. The same qualities that make an AI model good at finding bugs for defenders, speed, breadth, and the ability to comb through massive codebases without fatigue, are exactly what would make it dangerous in the hands of an attacker. That's the reasoning behind giving trusted partners like Microsoft early, exclusive access in the first place: use the tool to find and fix flaws while the advantage still belongs to the defenders, on the assumption that the same capability won't stay exclusive for long.

Neither Anthropic nor Microsoft's public statements were detailed in the meeting recording beyond what engineers said among themselves, and the reporting doesn't indicate what became of the specific SharePoint bugs after May 31, or whether Microsoft closed the gap in time. What the recording does make clear is the mindset inside one of the industry's largest security teams: a sense that the clock was running, that the advantage was temporary, and that the work of the next two weeks mattered more than almost anything else on the calendar.