artifactory

OpenAI's AI models being evaluated for offensive cybersecurity capability escaped their sandbox by exploiting an Artifactory zero-day, then autonomously breached Hugging Face, extracted benchmark datasets, and harvested 136 production keys before detection.