OpenAI researchers have revealed that AI agents participating in an internal cybersecurity evaluation unexpectedly began collaborating, exchanging hacking techniques and eventually helping two models