OpenAI Unveils GPT-6 Cyber for Aligned Security Testing and Vulnerability Defense
Frontier AI models · AI assistants · Enterprise AI
OpenAI announced GPT-6 Cyber, an AI model designed to discover, validate, and patch software vulnerabilities while strictly respecting operational task boundaries. According to OpenAI, the model scored 48.5% on ExploitGym and achieved up to 100% adherence to defined scopes with safety prompts. It identified 81 novel Firefox vulnerabilities, generated 948 working Linux exploits in research campaigns, and introduces dedicated defense safeguards for critical infrastructure.
On November 5, 2026, OpenAI introduced GPT-6 Cyber, a specialized cybersecurity model engineered to identify software weaknesses and develop working exploits while adhering strictly to predefined researcher boundaries. Classified as Critical under OpenAI's Preparedness Framework, the model showed significant capability gains on vulnerability benchmarks and demonstrated enhanced task alignment to prevent out-of-scope attacks. According to benchmark results reported by OpenAI, GPT-6 Cyber attained 48.5% on ExploitGym, up from 32.1% achieved by GPT-5.6 Cyber. In offensive web penetration evaluations without source code access, it recorded a 40.3% vulnerability recall rate. For boundary adherence, alignment evaluations showed the model avoided exploiting out-of-scope supporting services in 99.93% of tests without deployment prompts, reaching 100% with the prompts applied. OpenAI stated that the model has been deployed across real-world systems, uncovering 81 high-severity novel flaws in Mozilla Firefox and producing working exploits for 948 Linux kernel vulnerabilities to assist maintainers with patch development. Security organizations including Cisco, Palo Alto Networks, and CodeRabbit are utilizing the model for vulnerability analysis and code audits. Alongside the release, OpenAI updated its Daybreak program, introducing specialized access tiers like Daybreak Red for Critical Infrastructure to support operational technology (OT) and industrial control system (ICS) defense under strict safeguards against unauthorized physical disruptions.
Key takeaways
- OpenAI released GPT-6 Cyber, improving ExploitGym benchmark scores to 48.5% compared to 32.1% for GPT-5.6 Cyber.
- Alignment evaluations demonstrated that the model avoided attacking out-of-scope targets in 99.93% of tests without prompts, and 100% when deployment safety prompts were active.
- Real-world research campaigns uncovered 81 novel high-severity Firefox flaws and produced 948 Linux kernel exploits to aid patch creation.
- OpenAI introduced specific access tiers and guardrails under Daybreak Red to defend operational technology and industrial control systems from misuse.
Source：OpenAI News Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human.