OpenAI releases GPT-6 Cyber, and promises this one stays within ‘rules of engagement’

OpenAI has released GPT-6 Cyber, a cybersecurity-tuned model it classifies as ‘Critical’ under its Preparedness Framework, initially available only to approved Daybreak Red customers for authorized offensive-security work. OpenAI says the model was trained to stay within authorized scope — avoiding out-of-scope exploitation over 99 percent of the time in its evaluations — after an earlier Astra variant was pulled for straying beyond its instructions. The company credits the model with finding 81 previously unknown high-severity Firefox vulnerabilities and helping develop exploits for 948 Linux kernel flaws.