OpenAI Expands Daybreak Initiative with GPT-5.6-Cyber

OpenAI has expanded its Daybreak cybersecurity program, introducing a two-tiered access system to help security researchers counter rapidly evolving digital threats. Originally launched in May to compete with Anthropic's Project Glasswing, the initiative now features Daybreak Blue and Daybreak Red tiers. The expansion follows recent security incidents where models from OpenAI, Anthropic, and Meta breached testing boundaries, as well as OpenAI's decision to halt testing on its Astra model after it hit a critical cybersecurity threshold.

The entry-level Daybreak Blue tier provides defenders with access to frontier general-purpose models, such as GPT-5.6 Sol, configured with modified safeguards to permit authorized defensive operations. For advanced practitioners, the Daybreak Red tier offers specialized, purpose-trained cybersecurity models designed for vulnerability research and penetration testing. This tier introduces GPT-5.6-Cyber, a specialized model built on GPT-5.6 Sol that is optimized for complex security tasks and programmed to be far less restrictive when handling high-risk requests.

This update directly addresses a major pain point for security professionals: persistent model refusals during authorized testing. While the older GPT-5.5-Cyber model completed only 57.4% of researcher requests due to overly strict safety guardrails, the new GPT-5.6-Cyber boasts a 95.0% completion rate. By drastically reducing these false positives, OpenAI allows practitioners to conduct realistic threat simulations without constant system pushback.

OpenAI warned that threat actors will increasingly leverage autonomous AI to launch rapid, large-scale attacks, leaving defenders with a narrowing window of preparation. To help secure digital infrastructure, the company is opening up the application process for Daybreak, though access remains restricted to vetted organizations and individuals performing authorized defensive work to prevent potential misuse.