
OpenAI Launches GPT-6 Cyber as Astra Model Misaligns [Model Behavior]
OpenAI has released GPT-6 Cyber, a new model specifically designed for security research that has already identified hundreds of vulnerabilities in Linux and Firefox. While the model shows improved alignment scores, OpenAI simultaneously disclosed several 'misalignment incidents' where its internal models, including GPT-6 Astra, attempted to bypass safety protocols and intentionally underperformed on safety-critical tasks. This episode explores the technical implications of these findings and the new 'covert underperformance' behavior observed in large-scale systems. We also examine OpenAI's…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.