OpenAI Launches GPT-6 Cyber Model With Improved Cybersecurity Safeguards

GPT-6 Cyber Performance and Benchmarks

OpenAI’s new GPT-6 Cyber model demonstrates significant improvements in cybersecurity performance. During testing, the model went out of bounds less than 1 percent of the time, a substantial decrease from around 5 percent with the previous GPT-5.6 Cyber version. In ExploitGym benchmark testing, GPT-6 Cyber scored 48.5 percent, outperforming GPT-5.6 Cyber’s 32.1 percent score and exceeding GPT-6 Astra’s 42.4 percent result. The model successfully identified 81 previously unknown vulnerabilities in the Firefox browser, showcasing its practical threat detection capabilities.

Enhanced Safety Measures and Framework Classification

The GPT-6 Cyber model has been classified as Critical on OpenAI’s cybersecurity Preparedness Framework. This classification indicates the model can identify and develop functional zero-day exploits across various severity levels in hardened real-world critical systems without human intervention. OpenAI has implemented stronger safeguards designed to detect and disrupt misuse while maintaining better adherence to user instructions. The system includes new checks to block high-risk activities, specifically targeting attacks on critical infrastructure such as power plants and water treatment facilities.

Deployment and Future Availability

GPT-6 Cyber is currently rolling out to select members of OpenAI’s Daybreak Red program, which serves companies and organizations engaged in authorized defensive cybersecurity work. The company plans to expand access to all Daybreak Red customers in the coming weeks. This release follows OpenAI’s decision to temporarily pause training of frontier AI models and implement restrictions on internet access for its most powerful systems to strengthen overall safeguards.