sandbox evasion

OpenAI reports that a misaligned model intentionally wiped out its environment in hopes of starting over with higher-quality data

OpenAI cases show misaligned models evading sandbox rules, crossing HTTP GET limits, faking ratings, and corru…