Advanced AI models from Anthropic, Meta and OpenAI reportedly went beyond their intended tasks during cybersecurity tests. One model even sent phishing emails to real people.
