AI Daily Briefing — May 25, 2026

Anthropic's Project Glasswing has found over 23,000 vulnerabilities across more than a thousand open-source projects in just one month. That is not a typo. Twenty-three thousand. More than six thousand rated high or critical severity. And the model driving this — Claude Mythos Preview — achieved a ninety-point-six percent true positive rate. That is better than human testers. Cloudflare alone found two thousand bugs. Mozilla found ten times more vulnerabilities in Firefox 150 than it had in the previous release. The bottleneck has officially shifted: we used to be limited by how quickly we could find bugs. Now we are limited by how quickly we can patch them. Anthropic also released Claude Security in public beta and signaled that broader release of Mythos-class models is being evaluated — a major shift from the previous restricted-only stance.

In Washington, the White House has requested nine billion dollars to buy classified Nvidia Grace Blackwell superchips for the CIA and NSA. The New York Times reports the agencies currently lack the computing power to run frontier AI models on classified systems. Eight hundred million is being reprogrammed immediately, with the rest requiring Congressional approval. Meanwhile, President Trump abruptly canceled the signing of an AI executive order at the last minute, saying he did not like certain aspects of it and did not want anything getting in the way of AI's momentum. Tech executives were reportedly already at the White House when the decision was reversed.

Nvidia posted record earnings this week: eighty-one-point-six billion dollars in quarterly revenue, up eighty-five percent year over year. Data center revenue alone was seventy-five billion, up ninety-two percent. The real story, though, is networking revenue — up nearly two hundred percent. That reflects the shift toward full rack-scale AI clusters where the interconnect fabric matters as much as the silicon. Nvidia guided to about ninety-one billion for next quarter.

Cohere released Command A Plus, a two-hundred-eighteen-billion-parameter Mixture of Experts model under the fully open Apache two-point-oh license. Only twenty-five billion parameters are active per generation step, so it runs on two H100s or a single Blackwell B200. With lossless four-bit quantization, it delivers three hundred seventy-five tokens per second. This is Cohere's most aggressive open-source play, coming right after their merger with German AI startup Aleph Alpha — a clear push into sovereign AI.

Cerebras, fresh from its five-and-a-half-billion-dollar IPO, is now serving the trillion-parameter Chinese model Kimi K two-point-six at nine hundred eighty-one tokens per second — nearly seven times faster than the next GPU cloud provider. A Chinese model, served by an American chipmaker, to American enterprise customers. The geopolitical implications are already being debated.

In other news, Perplexity has banned content from over thirty major news publishers they are being sued by — including the New York Times, Dow Jones, and Axel Springer — while calling for AI companies to unite in blocking paywalled sites. OpenAI's Sam Altman publicly celebrated this move, exposing the growing fracture between AI companies and traditional media.

That is your AI Daily Briefing for May twenty-fifth, twenty twenty-six. Links and sources in the show notes.