What prevents the LLM from becoming an adversarial model to defeat the LLM that the company uses? Then you're turning the interview process into an arms race of

What prevents the LLM from becoming an adversarial model to defeat the LLM that the company uses? Then you're turning the interview process into an arms race of "Who has access to the best adversarial models?" and thereby make it completely redundant since at this point you're paying to get a job.