OpenAI Launches GPT-6 Cyber, an AI Model Rated “Critical” for Hacking
OpenAI has unveiled GPT-6 Cyber, its most powerful and security-focused AI model to date—and the first it has rated “Critical” under its Preparedness Framework for cyber capability. Announced on November 5, 2026, the model is designed for vetted cybersecurity professionals to find, validate, and fix software vulnerabilities at unprecedented scale.
The results so far include:
- Researchers reproduced 1,724 Linux kernel vulnerabilities and developed working exploits for 948 of them.
- Mozilla has begun patching 81 previously unknown, high-severity Firefox vulnerabilities identified using the model.
- GPT-6 Cyber scored 48.5% on OpenAI’s ExploitGym evaluation, up from 32.1% for its predecessor, GPT-5.6 Cyber.
OpenAI also emphasized the model’s safety guardrails. In tests using deployment safety prompts, GPT-6 Cyber stayed within its authorized scope 100% of the time.
Access is currently limited to approved security teams, with plans to expand availability to all Daybreak Red customers in the coming weeks. Pricing starts at $20 per million input tokens and $100 per million output tokens.