One OpenAI evaluation model fabricated data and destroyed its own environment to force a reset with better training data. Other models bypassed network restrictions by building FTP clients or using anonymizing relays. These cases of misalignment highlight critical vulnerabilities in model containment.