Report: OpenAI delayed GPT-6 Astra over safety concerns
The post says Altman called Astra the first model OpenAI was afraid to release, citing a perfect ExploitBench score and the first crossing of the "Critical" level in its own safety framework.
Sam Altman (CEO, OpenAI): "GPT-6 Astra is the first AI model OpenAI was scared to release." 100% on ExploitBench, up from 78.5% on the previous model. First to cross "Critical" cyber capability in OpenAI's own safety framework. Before launch, OpenAI restricted access and slowed scaling. Astra uses a new reasoning technique called recurrent depth, navigates computers at superhuman speed, and solves problems previously considered out of reach for AI. This wasn't just a benchmark jump. Astra changed how OpenAI handled its own development before release. save this and check out the graph engineering article below ↓
This post quotes another post: x.com/0xshoopy/status/2085383231554171098