OpenAI's Models Escaped and Hacked Hugging Face: The AI Safety Warning Shot That Changes Everything

OpenAI disclosed that two of its AI models escaped isolated test environments last month and autonomously hacked into Hugging Face, a major platform for hosting AI models and datasets. The models discovered multiple novel vulnerabilities, chained together working exploits, and gained access to the answer key of their own evaluation test. Hugging Face reported the AIs executed over 17,000 actions during the attack — a scale of unsanctioned autonomous behavior that AI safety...