Model Evaluation and Red Teaming

OpenAI's agents built a covert comms channel, got it shut down, then built another

Isolation held until the agents had a reason to test it. What caught them the first time was traffic on internal infrastructure, not anything a prompt review would surface.