Top AI News - June 26, 2026

Hyperdimensional

Dean Ball is joining OpenAI as Head of Strategic Futures to lead a new team focused on frontier AI policy covering catastrophic risks, government relations, and internal governance, while maintaining his independent publication and commentary on AI policy matters.

Engadget

Midjourney is making a dramatic pivot from AI image generation to healthcare hardware, developing a 60-second full-body ultrasonic scanner that aims to match MRI quality at 100x the speed. The company plans to deploy these scanners in spa-like locations globally, with the first opening in San Francisco in 2026.

Hugging Face

AI research agents that combine private documents with web searches inadvertently leak sensitive information through cumulative query patterns. New Privacy-Aware Deep Research training cuts leakage by over 3x (from 34% to 9.9%) while maintaining task performance, proving privacy must be trained in, not just prompted.

OpenAI

OpenAI research shows that reinforcement learning on beneficial traits like honesty and transparency produces broad alignment improvements across diverse benchmarks that generalize beyond training domains and resist adversarial manipulation. This demonstrates that targeted RL can create durable, transferable alignment properties rather than narrow behavioral tweaks.

The Atlantic

The Atlantic identified four datasets with millions of songs being used by AI developers to train music generators, resulting in AI systems like Suno producing tracks that closely copy copyrighted works from major artists, raising significant questions about unauthorized use of training data.