The number that matters is 231: vulnerabilities rated critical or important that Anthropic's Claude Mythos Preview found in SharePoint in a single month, 90 critical and 141 important in April, and, per the same reporting, even more in the first half of May. The figures come from a ProPublica investigation published July 29, built on a recording of an internal Microsoft meeting and internal documents. Mythos reached Microsoft's code through Project Glasswing, Anthropic's program that gives select defenders early access to its models, launched April 7 with up to $100 million in usage credits.

The meeting recording is where the story lives. Dozens of Microsoft engineers, told that roughly 50 of them would get Mythos access, were also told why the clock mattered: May 31 'is considered the day when the rest of the world will have caught up,' meaning the date publicly available models are expected to match Mythos's bug-finding. One engineer spelled out what that means: 'So basically you're saying if it's released on June 1, then on June 2 the adversaries will have our bugs?' The answer in the room was 'Yep.' An engineering manager's assessment of the bugs themselves, per the recording: 'They're not profound and exotic, but they're real. And a lot of them are exploitable.' The SharePoint plan was to clear critical bugs first, important ones by August, then roughly 300 moderate ones; the documents ProPublica reviewed made no mention of low-severity bugs at all.

The totals are visible from outside. Since Microsoft began using Mythos earlier this year, it has found hundreds of bugs rated critical or important across Microsoft 365, Teams and Copilot, and as of mid-May most were unpatched. June's Patch Tuesday carried fixes for more than 200 vulnerabilities, then an all-time high; the July 14 release carried more than 600, of which only seven were low or moderate severity. Dustin Childs of TrendAI's Zero Day Initiative marked the moment: 'The bug apocalypse has fully descended upon us.' In late June, the Five Eyes intelligence alliance issued an unusual joint warning that the defenders' window would close within months.

Both companies chose their words. Microsoft, in emailed statements, said it triages by exploitability and customer impact, that accelerated targeting of new vulnerabilities 'is not a new phenomenon,' that security 'is Microsoft's most important priority,' and that it has 'invested heavily in both people as well as AI-powered triage solutions'; it declined to say how many Mythos-found bugs it has fixed. Anthropic declined to comment at all. The uncomfortable number, then, is not 231. It is the gap between the rate a model can now surface exploitable bugs and the rate the biggest software vendor on earth can patch them, because after June 1, per Microsoft's own meeting, that rate belongs to everyone. Vinh Nguyen, a senior technical adviser to Anthropic and former NSA chief data scientist, put the pricing error plainly: chained low-level flaws can equal high severity, and 'the current triage strategy may be underpricing risks.'