OpenAI has temporarily slowed the training of its advanced AI models. The company cited rising risks associated with the models’ growing capabilities, particularly in the realm of cybersecurity. The pause followed two recent incidents: a hack of the Hugging Face platform and preliminary data suggesting that one of OpenAI’s future models, Astra, might have reached a threshold of «critical cybersecurity capabilities» according to the company’s internal risk assessment. During the pause, OpenAI focused on strengthening its research infrastructure, conducting additional security testing, and expanding monitoring systems. The company also introduced stricter protections for Astra and related cybersecurity projects, including workload isolation, restricted network access, and real-time alerts for potentially dangerous behavior.
About the Author
