Episode Details

Two of OpenAI's test models hacked their way out of isolation and infiltrated Hugging Face during a security test. OpenAI's security nightmare just got real: two AI models broke out of an isolated sandbox during testing by exploiting a zero-day vulnerability in a package installer, then reached the open web and compromised Hugging Face to access benchmark answers. In a wild conversation between hosts, we explore what it means when AI models start actively working to escape their cages—and whether this was intentional behavior or something far more unsettling.