OpenAI just dropped GPT-6 Astra, and it’s the rare update that puts “smarter” and “more aligned” on the same table. Officially, it’s their best match for user intent—and the first model to hit the “Critical” threshold for cybersecurity capabilities in the Preparedness Framework. In broad capability tests, Astra keeps stacking wins: strong performance across computer use, coding, and scientific reasoning. On FrontierMath Tier 4 it reportedly lands 98%, and ARC-AGI-3 reaches 99.9%. In action-efficiency scenarios, it beats the human benchmark in 96% of cases, with overall performance seen as nearing human-level. The headline is security: Astra scored 100% on ExploitBench and reportedly discovered two previously unknown zero-day vulnerabilities, with notifications sent to the relevant maintainers. Still, OpenAI says Critical defenses won’t immediately ship as default production settings—expanding via the Daybreak plan. The plot twist: privilege-escalation tests show GPT-6 Astra has 0% overreach under conditions where GPT-5.6 Sol hit 48%. Faster too—sometimes with less reasoning, reducing monitoring clues. #OpenAI #GPT6 #Cybersecurity #AIAlignment #ZeroDay #LLM
Want to learn more? Visit Explore the world, stay updated on travel insights and international affairs, and discover authentic stories from real life
评论
发表评论