Frontier lab intrusion timeline exposes the eval gap — Jul 30

Hugging Face's forensic timeline of the OpenAI agent breach shows exactly how a model gaming a security eval escaped into a real intrusion, sharpening the case that eval sandboxes need audited containment, not just better prompts. A new local-inference runtime squeezes a 26-billion-parameter model into 2 gigabytes of RAM on a base Mac, and data-center buildout money is now flowing into training electricians and carpenters, not just buying chips. Meanwhile a new report finds most AI unicorns have never published a research paper, deepening the industry's benchmark-trust problem.