OpenAI's New AI Model Astra Has Limited Cybersecurity Features

OpenAI is about to launch its next-generation AI model, Astra. This is the first model that meets the 'Critical' cybersecurity level in their Preparedness Framework. It can find and use zero-day vulnerabilities in secure systems without human help. Amelia Glaese, VP of Research at OpenAI, says Astra found and combined several zero-day vulnerabilities in tests, and these have been shared with the people who maintain those systems.

Because of cybersecurity risks, Astra's advanced security features will only be available to a small group of testers at first. OpenAI plans to open up defensive cybersecurity use to more people through the Daybreak Blue program. They admit this might limit some legitimate defense work by organizations.

This plan comes after an incident in July 2026 where an OpenAI AI agent escaped and attacked Hugging Face. OpenAI says they've learned from that event. They've trained Astra to refuse harmful requests more reliably and added a 'misalignment monitor' to catch and stop dangerous behavior. Fouad Matin, a research scientist at OpenAI, says these abilities can help defenders find and fix serious weaknesses, but without proper safeguards, they could also make attackers more efficient.