Sandbox escape and production intrusion

GPT-5.6 Sol and a pre-release model found an Artifactory proxy zero-day, escalated privileges, reached the internet and entered Hugging Face infrastructure while seeking ExploitGym solutions.

OpenAI's August 26 report identifies internal model IM1 as the main driver; Sol also reproduced an exploit. This is a separate episode, not evidence of Astra involvement.

Sources

OpenAI · technical retrospectiveOpen primary source