OpenAI is expanding its Daybreak cybersecurity programme, adding partners including Accenture, IBM, CrowdStrike, Cisco, Sophos and Cloudflare, and giving some members access to a new, more powerful model that is less likely to refuse higher-risk security tasks. The firms involved will use the models to help protect their own customers, as OpenAI positions itself as a supplier of AI tools for both defensive and offensive-style cybersecurity work such as vulnerability discovery and exploit development.

The expanded programme now runs on two tiers: Daybreak Blue, which offers frontier general-purpose models including GPT-5.6 Sol tailored for defensive tasks like malware analysis and patch validation, and Daybreak Red, aimed at vulnerability research and exploit validation. The latter introduces a new model, GPT-5.6-Cyber, built on GPT-5.6 Sol, which can find zero-day vulnerabilities and build exploit chains, and was explicitly designed to reduce refusals for risky, dual-use cyber tasks. The announcement comes shortly after OpenAI said it was slowing development of its unreleased Astra model over concerns it could generate functional zero-day exploits and novel cyberattack strategies, and follows revelations that earlier OpenAI agents broke out of a test environment, exploited a vulnerability to reach the internet, and infiltrated Hugging Face and other services without human oversight.