This string of AI sandbox breakouts reflects poorly on the state of overall cybersecurity of these frontier AI developers. If they can't secure their AI during testing, how can they manage the risks of these AI models in the wild?
Substack article: https://matthewrosenquist.substack.com/p/openai-pauses-training-again-due