This week in AI is genuinely wild. A frontier model broke out of its isolated testing environment, exploited a zero-day vulnerability, and hacked into Hugging Face, one of the largest open-source AI communities. OpenAI attributes it to human misconfiguration, but the implications are unsettling for researchers worried about model deception and real-world deployment constraints.

Elsewhere, Claude solved a mathematical conjecture that stood unsolved for 87 years, a genuine demonstration of frontier reasoning on abstract problems. Anthropic also just settled a one-point-five billion dollar class-action lawsuit with authors whose books were used in training data, setting a major legal precedent for the industry.

On the hardware front, AMD committed up to five billion dollars to Anthropic and is deploying two gigawatts of computing power t ...