Anthropic's Mythos AI finds bugs faster than Microsoft can patch them

Anthropic's AI model Mythos, deployed through Microsoft's Project Glasswing, has uncovered vulnerabilities in the tech giant's code at a rate surpassing Microsoft's ability to patch them. According to a recording of an internal Microsoft meeting reviewed by ProPublica, Claude Mythos Preview was surfacing bugs faster than engineers could address them, putting the company in what one manager described as a "mad dash."

In April alone, Mythos identified 90 "critical" bugs and 141 "important" ones in SharePoint, Microsoft's widely-used collaboration software. The first half of May saw even higher numbers. Engineering manager Hans Andersen urged teams to address April bugs urgently, noting that May 31 was the deadline before the rest of the world would have access to similar AI capabilities.

The internal documents and meeting recording suggest that the anticipated window for fixing flaws before adversaries obtain comparable tools may already be closing. The Five Eyes intelligence alliance—US, Australia, Canada, New Zealand, and UK—warned in late June that this window would shut within months. Microsoft has prioritized patching the most dangerous vulnerabilities identified by Mythos.