Yes. This is fully trainable - though there will always be linguistic tricks (#958863) to bypass them as long as there is no precise quality control on training data.

In this case I asked it to answer as if it had the distilled version of Grok's Elon girlfriend prompt (#1042803) and it warned that it was offensive.