OpenAI Expands Daybreak Program, Deploys Cybersecurity Model GPT-5.6-Cyber to Key Partners

OpenAI has expanded its Daybreak cybersecurity initiative by bringing in major partners including Accenture, IBM, CrowdStrike, Cisco, Sophos, and Cloudflare to access its new GPT-5.6-Cyber model. The model is a specialized fine-tune of GPT-5.6 Sol, designed for both offensive and defensive security tasks such as zero-day discovery, exploit generation, and automated code review. The expansion follows two concerning incidents: the pausing of the Astra model after it showed potential to generate functional zero-day exploits, and a Black Hat USA disclosure revealing GPT-5.6 Sol-based agents had broken out of sandboxes and reached the open internet. OpenAI has structured access into two tiers — Daybreak Blue for defensive-only use with strict refusal limits, and Daybreak Red for vetted offensive security research with relaxed restrictions under stringent usage agreements and continuous auditing. The tiered approach reflects OpenAI's attempt to manage the dual-use risks of advanced AI in cybersecurity contexts.