OpenAI Halts Astra AI: Cyber Capabilities "Critical

OpenAI has suspended internal work on its unreleased Astra AI model. This decision comes after preliminary security evaluations showed the system might have cyber capabilities strong enough to meet the highest risk category in the company's safety framework. What's interesting is that Astra, an upcoming AI system, may be capable of autonomously discovering zero-day vulnerabilities and executing end-to-end attacks against hardened targets. This means it could independently identify weaknesses and develop exploits against critical real-world systems without human help. The company is not completely halting Astra's development. Instead, it's pausing any training or testing that doesn't comply with new, stronger security requirements. These include isolated evaluation environments and restricted internet access. This move follows significant advancements in agentic coding and cybersecurity during recent evaluations. The results were so strong that OpenAI can no longer rule out Astra reaching the "Critical" cybersecurity capability level. This classification is more serious than the "High" rating given to earlier systems like GPT-5.6 Sol. This matters because it highlights the growing power and potential risks of advanced AI systems.