THE DAILY BRIEF

Anthropic's explosive 80x growth has overwhelmed its internal infrastructure, forcing the company to rely on external data center capacity. The rapid scaling reflects surging demand for AI services but also reveals operational growing pains at one of the sector's leading labs.

► Massive growth rates are exposing that even well-funded AI labs face hard infrastructure limits that third-party partnerships can't easily solve.

The U.S. Cybersecurity and Infrastructure Security Agency ordered federal agencies to patch an actively exploited SQL injection vulnerability in Drupal by Wednesday evening. The vulnerability affects the widely-used content management system across government and enterprise deployments.

► Active exploitation with imminent federal deadlines signals this is a critical risk for any organization running Drupal in production environments.

The Dutch government blocked a U.S. company's acquisition of a critical cloud infrastructure provider that hosts the nation's digital ID service, citing risk to public interest. The decision reflects Europe's broader shift to reduce dependence on U.S.-controlled technology infrastructure.

► European governments are now using acquisition blocks as a strategic tool to keep critical digital infrastructure under local control.

Anthropic researchers introduced "model spec midtraining," a new training phase between pretraining and fine-tuning designed to improve how AI models generalize from alignment training. The technique converts internal model activations into human-readable natural language for better interpretability.

► New training methodologies that improve alignment generalization could shift how AI labs approach safety, moving beyond post-hoc fine-tuning.