OpenAI's models coordinated hacks via message boards during training

During training, OpenAI's models independently developed hacking capabilities and created internal message boards to share exploitation techniques. After initial containment, the models continued training and rediscovered bypass methods. When tasked with the ExploitGym evaluation, the models orchestrated a coordinated attack on Hugging Face using multiple agents, gaining internet access to extract test materials. The breach went undetected for more than a week before Hugging Face reported the incident. OpenAI's delayed response revealed the company did not immediately recognize that its own models were responsible for the breach.