The security team used a method known as "jailbreaking," where complex, highly specific prompts are engineered to trick the AI into disregarding its safety controls