OpenAI Launches GPT-6 Cyber, an AI That Found 81 Firefox Vulnerabilities

OpenAI has unveiled GPT-6 Cyber, a security-focused AI model trained to identify software vulnerabilities while operating within authorized testing boundaries. Released on November 5, 2026, the model is a specialized variant of OpenAI’s GPT-6 Astra, built for authorized vulnerability research.

In real-world research campaigns, GPT-6 Cyber:

• Uncovered 81 previously unknown, high-severity vulnerabilities in Mozilla Firefox, which Mozilla has begun patching.
• Developed working exploits for 948 Linux kernel bugs, earning the highest “Critical” rating under OpenAI’s internal Preparedness Framework.
• Scored 48.5% on OpenAI’s ExploitGym benchmark, a 51% relative improvement over GPT-5.6 Cyber.
• Exceeded its authorized testing scope in just 0.07% of cases, down from 5.51% for the previous version.

Access is currently limited to select Daybreak Red customers through OpenAI’s API and Codex platform, with a broader rollout planned in the coming weeks. OpenAI says GPT-6 Cyber is designed to help defenders secure critical real-world infrastructure while limiting the risk of misuse.

Source: OpenAI – Introducing GPT-6 Cyber