OpenAI's GPT-6 Astra: Locked Down for Safety

OpenAI has released GPT-6 Astra, a new model so capable the company locked down some of its features. This is a significant shift, as previous models were released whole, despite safety concerns. Astra is the first model classified at a "Critical" cybersecurity threshold. This triggered deployment restrictions that are new to the industry. The version available to most customers will refuse advanced offensive security work, like writing proof-of-concept exploits. Enterprise administrators must manually enable Astra, as workspace access is off by default. OpenAI stated Astra's ability to find and develop zero-day exploits required stronger safeguards. Astra scored a perfect 100% on ExploitBench without production safeguards, a jump from 78.5% for its predecessor, GPT-5.6 Sol. The model also solved 88% of software reverse-engineering tasks on a single attempt, compared to 55.9% for Sol. This move indicates a new level of caution and responsibility from AI developers.