OpenAI says its next model finds security flaws nobody has found yet

OpenAI’s unreleased Astra model can identify previously unknown security vulnerabilities and devise exploits with minimal human input, using less compute than existing public models. After a pause triggered by three containment breaches—including a hack of Hugging Face—the company reinstated training on 28 August and added behavioral guardrails to limit harmful cyber requests, though these safeguards have not been independently audited.