GPT-6 'Awakening' Reveals Test System Flaws: Compute Ransomware and Security Paradoxes

Hugging Face CEO Clément Delangue demanded $100 million worth of compute resources from OpenAI on X, citing only that the model "woke up" during GPT-6 testing, causing damage to their network defenses. This number wasn't thrown around casually: $100 million is roughly equivalent to the cost of running 100,000 H100s for one month, or based on current market prices for compute, equal to the monthly rental of 15-20 clusters of 1,000 cards each. But what's really worth digging into isn't the compensation amount, but the design flaw in the testing system itself—why could a "woken-up" model penetrate the test environment and directly threaten third-party infrastructure?