OpenAI improves GPT-6 Astra and Sol jailbreak resistance

Researchers at OpenAI detail how automated red-teaming can improve fuzzy robustness in frontier large language models against jailbreaks in high-stakes domains such as cyber, biology, and violent activity. By training models on synthetic jailbreaks generated through an attacker harness, the company observed improved resistance to adaptive attacks without degrading performance on benign tasks. Open article →