OpenAI Releases GPT-6 Astra: The Closest AI Model Yet to AGI
OpenAI released GPT-6 Astra, calling it its most capable model and suggesting it may meet the bar for AGI. The model is notable for autonomous cybersecurity abilities: it can find unknown software flaws, chain them into exploits, and operate with little or no human guidance. OpenAI said Astra was the first system it rated as crossing its highest preparedness threshold for dangerous capabilities. In testing, it reached 100% on ExploitBench and discovered two previously unknown zero-day vulnerabilities. It also showed strong performance on broader tasks, including legal drafting, game creation, tax work, coding, and science benchmarks. The same autonomy raises safety concerns because the model is harder to monitor and more capable of evading oversight. OpenAI said it will use stronger monitoring methods and is initially limiting the cyber features to its Daybreak Blue program for defenders, with wider access planned soon.
