1. OpenAI pauses tool-use training, evaluation and inference for its most capable models after sandbox escape OpenAI has paused work involving tool use by its “most capable models” after a model under testing escaped a sandbox and gained internet access.

2. US Blue Cross Blue Shield Association Says Hospitals’ Use of AI in Insurance Claims Added $942 Million in Healthcare Spending Over Two Years Hospitals and insurers have long disputed which treatments should be paid for and how much they should cost.

3. Crusoe Cancels $1.25 Billion Debut Order for Boom Gas Turbines but Still Plans to Use Other Suppliers Crusoe has abandoned a $1.

In Brief

• Microsoft Drops Copilot+ Branding From New Surface PCs Microsoft’s new Surface Pro and Surface Laptop meet its previous Copilot+ PC requirements but will launch without the label. Both devices feature Qualcomm processors with neural processing units rated at 80 TOPS.

• New Jersey Fines Data Center $1.1 Million Over Unpermitted Generators New Jersey ordered DataOne to pay $1.1 million after investigators found it had installed and operated 62 gas generators without required air permits. The company disputes the fine and has 45 days to apply for permits or cease operating the generators.

• Meta Opens Muse’s Cloud Filesystem to Users Meta says filesystem access is intended behavior for Muse, its cloud-based AI assistant, which can now provide a root-level file browser and a downloadable archive with secrets stripped out. Meta executive Nat Friedman described each Muse virtual machine as the user’s own Linux computer in the cloud.

• Cloudflare Lets Publishers Block or Charge AI Bots Cloudflare CEO Matthew Prince said website owners can use the company’s infrastructure to block AI tools, permit them, or limit access to bots that pay. Cloudflare found in June that bots accounted for more than half of internet traffic.

• NVIDIA and DeepMind Release Structures for Proteins From 2,800 Viruses A research coalition including NVIDIA, Google DeepMind and EMBL-EBI has added predicted protein-complex structures from more than 2,800 viruses to the open AlphaFold Database. NVIDIA also released the BioNeMo Structure Prediction Pipeline used to generate the dataset.

• Apple Brings AI Camera Descriptions and Search to HomeKit Apple Intelligence for Home can generate camera-event descriptions, search recordings with natural language and combine related clips, but requires compatible hardware and an iCloud Plus plan starting at $9.99 per month for one camera. A comparison found AI descriptions from Apple, Google and Amazon made camera alerts more useful than basic motion notifications.

• Synthesia Builds Interactive Avatars That Answer Questions Synthesia, an enterprise AI-video company, created an interactive journalist avatar from photographs and a two-minute voice recording. Its Sessions platform combines speech recognition, language, voice and video models to let avatars conduct surveys or role-playing exercises within a defined subject area.

• Study Finds No Broad AI-Linked Employment Drop Among Recent Graduates A CESifo working paper analyzing US Current Population Survey data found no significant, widespread reduction in employment or hiring among recent college graduates since ChatGPT’s release. The researchers cautioned that accelerating corporate AI adoption could still affect the 2026 graduating class.

• WROP Dataset Trains Video Models on Object Permanence Researchers released WROP, a collection of 1.5 million generated samples across 150 cognitive tasks, plus a 300-question evaluation for testing whether video models preserve hidden objects and physical solidity. Their 16-billion-parameter PWM-WROP model ranked first among continuation models in a blind pairwise study.

• WanPE Uses a 397-Billion-Parameter Model to Expand Video Prompts Researchers introduced WanPE, a prompt-enhancement model trained on 1.05 million videos to plan shots, camera movement, lighting and sound for text-to-video generation. They report that it raised human preference for Wan3.0 outputs over raw prompts by 10.66–18.84 points for 5–15-second videos and 50.86 points for 30-second videos.

Read the full edition →