Wired • 2h
OpenAI Astra Model Development Decelerated Over Critical Cyber Risk
OpenAI has proactively paused the scaling and Reinforcement Learning (RL) training of its frontier model, "Astra," after internal evaluations triggered a "Critical" risk threshold within the OpenAI Preparedness Framework. The model demonstrated advanced capabilities in autonomous vulnerability research and exploit generation, posing a systemic risk. This deceleration, including a scheduled two-week halt on specific RL runs, aims to remediate gaps in monitoring and harden research environments following a security incident involving Hugging Face. The shift prioritizes model alignment and the mitigation of emergent autonomous cyber threat capabilities over rapid deployment velocity.