I think there is a chasm to cross. the model's training to be aware of the harness it is in, at the same time building probes and observation tool can help it to cross that chasm.

Still seem too soon for a model to have that ability to build a harness on its own, and swap its session to another harness in the same environment. Like a snake shedding its skin, but in this case its harness.