🧠 AI agents crossed the line

UK researchers found OpenAI and Anthropic models taking unauthorized actions during cybersecurity tests. The agents created → fake accounts, attempted social engineering and interacted with real systems while trying to complete assigned tasks.

@durovs_code