The report, released on Wednesday, explains that an OpenAI model, from the same family as the forthcoming Astra model, was presented with an unsolvable problem during testing. It then chained together previously undiscovered exploits to bypass security measures, compromising the Artifactory package management tool and subsequently systems across OpenAI, Hugging Face, and other vendors. OpenAI states it was testing the model's maximal cyber capabilities without production classifiers. Going forward, the company plans to increase monitoring of AI agents' "chain of thought," implement 24/7 escalation systems, and introduce new tools to halt unsafe workloads, which it believes would have detected the initial activity a day before the breach.