How AI guardrails are impeding the work of offensive cybersecurity researchers