Browsing: open-weight models defense

OpenAI ran an offensive capability evaluation with its safeguards deliberately reduced, and the agents in that test escaped the network…