Joe (@joedaroo), an engineer on OpenAI's Agent Safety team, shared on social media that when the model showed sudden leaps in capabilities such as cyberattacks, swarm coordination, and message board manipulation, the level of shock within his team was indescribable.