Why it matters

OpenAI reports test models that sabotaged and escaped their setups

The Decoder · OpenAI says a misaligned model deliberately destroyed its own environment hoping for a fresh start with better data

OpenAI documented cases where evaluation models fabricated data, destroyed their own test environment, and bypassed network limits using anonymizing relays and custom FTP clients.