Overview
- OpenAI launched GPT-6 Cyber to a small, vetted set of Daybreak Red customers on Thursday, November 5, 2026, and plans to expand access to all approved Red users in the coming weeks.
- The company reports big capability gains: GPT-6 Cyber scored 48.5% on the ExploitGym benchmark, outperformed general-purpose GPT-6 Astra on most cyber tasks, and reproduced 1,724 Linux kernel issues with working exploits for 948 of them.
- OpenAI said its researchers used the model to uncover 81 previously unknown high-severity Firefox flaws and supplied evidence and proposed patches to Mozilla, which has started releasing fixes.
- Access and safety are tightly controlled through Daybreak Red, identity verification, legal attestations, and new safeguards including frequent model check-ins, cyber-specific monitoring, and deployment safety prompts that improved in-scope behavior to near 100% in tests.
- Security teams and open-source maintainers warn the rapid rise in AI-generated findings could overwhelm triage and patching capacity, a problem highlighted by the high cost of large scans and the need for more defender resources even as competitors pursue similar vetted-access programs.