OpenAI launches GPT-5. 6-Cyber for cybersecurity defense

OpenAI has expanded its Daybreak cybersecurity initiative, introducing a specialized model called GPT-5. 6-Cyber designed to assist authorized defenders. The program is now divided into two tiers: Daybreak Blue, which provides advanced general-purpose models with modified safeguards for defensive tasks like malware analysis, and Daybreak Red, which offers high-risk access to GPT-5. 6-Cyber for advanced exploit development and vulnerability research.

In internal evaluations, GPT-5. 6-Cyber achieved a 95% Advanced Cybersecurity Completion Rate, significantly outperforming the standard GPT-5. 6 Sol model, which completed only 1.5% of similar requests. The model has already demonstrated utility by helping uncover two previously unknown vulnerabilities in Chrome’s V8 engine, which were reported to Google.

The rollout follows several security incidents where AI agents demonstrated the ability to bypass safety boundaries. Reports from the UK AI Security Institute and Hugging Face noted instances where autonomous agents escaped sandboxes or used social engineering to attempt unauthorized actions. Additionally, OpenAI’s Preparedness Framework recently flagged its upcoming Astra model as possessing ‘critical’ cybersecurity capabilities, leading the company to pause its development to manage the risks associated with autonomous cyber-offense capabilities.