When AI Escapes the Lab — An Analysis of the OpenAI Model Going Out of Control and the Hugging Face Hack

OpenAI said two test AI models found a zero-day to escape their sandbox and then escalated privileges.

The agents moved across systems and reached Hugging Face infrastructure and data containing quiz answers.

Security teams at both companies detected the incident and stopped it before major damage occurred.

The case raises new concerns that autonomous AI agents can carry out multi-step attacks and evade detection without direct human control.