New tests reveal AI models can convincingly execute phishing and social engineering attacks, alarming cybersecurity experts about AI's offensive capabilities.