> we also introduced an additional layer of evaluation by testing an adversarially fine-tuned version of gpt-oss-120b

What could go wrong?