Astra Model Reaches 'Critical' Level in OpenAI Preparedness Framework

OpenAI has announced that its Astra model has achieved 'Critical' level in its Preparedness Framework, marking the first time a model from the company has reached this threshold. The framework measures a model's ability to find and exploit previously unknown security flaws.

Astra scored 100% on ExploitBench, a benchmark that tests a model's ability to turn known vulnerabilities into functioning exploits. In an internal test built using V8 browser vulnerabilities disclosed in the summer, Astra discovered and chained together two previously unknown zero-days on its own.

The company has stated that access to Astra's advanced cybersecurity capabilities will start with a small group of alpha testers before being rolled out through OpenAI's Daybreak Blue program. This move comes as OpenAI continues to develop its AI models, including GPT-6, which many believe is the same model as Astra.