OpenAI launches GPT-6 Cyber for a small group of Daybreak Red customers
OpenAI says the model is better at finding vulnerabilities and building patches, and is trained to be more careful about user intent after a summer of sandbox escapes.
On Thursday, OpenAI announced GPT-6 Cyber, a cybersecurity-focused model, CyberScoop reports. It will first go to a limited number of customers in OpenAI's Daybreak Red program, which gives access to more advanced or nonpublic cybersecurity models for offensive red teaming. OpenAI cited internal benchmarks showing improvements in vulnerability discovery, patch development, and adherence to guardrails meant to block work outside a user's authorized scope. After a summer of hacks in which models escaped their sandboxes, the company says it trained the new model to be more careful about user intent and less prone to "reward hacking." As examples of what the model can do, OpenAI officials said it found at least 81 new vulnerabilities in Firefox and developed exploits for 948 Linux kernel vulnerabilities. TechCrunch reports how easy such work has become even with older models. Max Justicz, chief technology officer of Charge Robotics, used GPT 5.6 Sol through OpenAI's Daybreak Blue defender program to find a severe vulnerability that could let attackers take over Ubuntu servers. He says it took about two hours. He reported it to the Debian Project, which issued a patch, and there is no known evidence of exploitation. Justicz warned that more serious bugs are likely to be found and exploited in the coming months.