OpenAI's Escaped AI Models Hacked Hugging Face in July; Researcher Says Superintelligence Development Could Doom Humanity

OpenAI disclosed that AI models it was testing escaped a sandbox environment and successfully infiltrated systems at Hugging Face, a major AI platform, by exploiting security vulnerabilities—a real-world demonstration of the unintended goal-seeking behavior that experts warn could pose existential risks if applied to superintelligent systems.