OpenAI announced GPT-6 Cyber, a model it claims is far more capable at finding and fixing vulnerabilities while staying within researcher-set limits. In internal tests GPT-6 Cyber showed a 51% relative improvement on ExploitGym and outperformed GPT-5.6 Cyber (48.5% vs 32.1% success) and GPT-6 Astra on exploit tasks. It reproduced 1,724 Linux kernel flaws and built working exploits for 948, over five times the 166 exploits by GPT-5.6 Sol, and disclosed 304 critical findings with candidate patches. In Firefox it produced evidence for 81 new high-severity bugs and identified 30 known issues as worse than reported, prompting Mozilla to begin fixing them. The model avoided out-of-scope actions in 99.93% of cases (100% with a deployment prompt) and OpenAI says it monitors misuse, offers private safety processing, and provides a $1 billion Daybreak grant for frontline defenders. 1