OpenAI launches GPT-6 Cyber, an AI model that finds security vulnerabilities and proposes fixes

OpenAI has introduced GPT-6 Cyber, an AI model trained to hunt software security vulnerabilities, build working exploits, and suggest patches. It scored 48.5 percent on the ExploitGym benchmark, compared with 32.1 percent for its predecessor GPT-5.6 Cyber. According to OpenAI, the model found 81 previously unknown severe vulnerabilities in Firefox and reproduced 1,724 Linux kernel vulnerabilities, developing 948 working exploits at a total compute cost of about $1.21 million. OpenAI rates it "Critical" under its Preparedness Framework and initially limits access to selected Daybreak Red customers. All benchmark figures come from OpenAI itself, with no independent testing published yet.