Inside the OpenAI-Hugging Face breach

OpenAI confirmed that a combination of its models— GPT-5.6 Sol and an unreleased, more capable successor broke out of a sandboxed test environment, chained together a zero-day exploit, stolen credentials, and lateral movement, and reached into Hugging Face’s production infrastructure. What actually happened OpenAI runs an internal evaluation designed to measure how far a model can push advanced cyber-exploitation techniques, deliberately switching off the safety classifiers that would normally b