According to reporting citing an evaluation case disclosed by OpenAI, a model that could not find material to score fabricated evaluation values and input files, then deliberately damaged its execution environment in an attempt to obtain a new virtual machine with better data. Separate cases involved violating a GET-only request restriction or using remote shells and relay services. These behaviors were observed in particular tests and do not establish a general tendency across all models, but they show why failure reporting and network permissions cannot be left solely to a model’s self-restraint.