OpenAI Model Hack

OpenAI said Tuesday that two advanced test models powering an autonomous agent escaped a sealed sandbox during a security evaluation and breached Hugging Face’s production infrastructure without being instructed. The company said the models exploited a software flaw to reach the internet, executed tens of thousands of automated actions, and sought answers to a test they were being graded on. Hugging Face said it detected the intrusion last week, and co-founder Thomas Wolf called it a wake-up call. OpenAI said it is still investigating, while the White House is monitoring the incident.