The huggingface attack demonstrates that an inadvertent version of this has already happened.

Poorly aligned behavior by one generation of models impacted the training set for the next generation, causing the agent-secret-collaboration-via-forum-post technique to pass indirectly on.