Offensive security researchers downgrade proprietary American commercial models to bypass increasingly restrictive artificial intelligence safety guardrails