I've heard that it was possible to trigger really obvious output poisoning on Fable with something as basic as asking the model to think outside of its built-in hidden thinking delimiters.

This watermark may trigger a similar mechanism.