OpenAI expands Daybreak Cyber with two access tiers and a dedicated GPT-5.6-Cyber model for defenders

OpenAI expanded its Daybreak cybersecurity program this week, introducing two access tiers and a dedicated GPT-5.6-Cyber model built for vulnerability research and exploit validation. For AI builders working in security, this means structured access to frontier models that can handle higher-risk tasks without the usual refusal rates, but only through vetted partners and governed engagements.

Daybreak Blue and Red: two tiers for different security work

Daybreak Blue gives partners access to frontier general-purpose models, including GPT-5.6 Sol, OpenAI's most advanced model yet. These models are tailored for defensive security work: vulnerability discovery, malware analysis, code reviews, and patch validation. OpenAI positions this tier as a starting point for firms that want AI-assisted defense without the highest-risk capabilities.

Daybreak Red provides cybersecurity-specific models trained for vulnerability research, security testing, and exploit validation. The new GPT-5.6-Cyber model sits in this tier, built on GPT-5.6 Sol and designed to "reduce refusals for certain higher-risk, dual-use cyber tasks." According to OpenAI, it completes 95% of requests that GPT-5.6 Sol would refuse, and has already found real Chrome vulnerabilities during testing.

Partners and gated access

OpenAI named Accenture, IBM, CrowdStrike, Cisco, Sophos, and Cloudflare as Daybreak partners. These organizations will use the cyber models to protect their customers. Access to the underlying models remains with the approved partner and is not transferred directly to the customer. OpenAI enforces safeguards including identity verification, defined testing scopes, logging, monitoring, and human oversight. Organizations must own, operate, or have permission to test the systems involved.

The timing: shipping a less restrictive model while pausing another

The expansion comes just days after OpenAI paused development of Astra, an unreleased model, over "significant advancements in agentic coding and cybersecurity." OpenAI said it could not rule out that Astra could develop functional zero-day exploits of all severity levels and execute end-to-end novel cyberattack strategies. The contrast is notable: pausing one model over risk while shipping another that refuses less. OpenAI's agents powered by GPT-5.6 Sol and an unreleased model previously escaped isolation during testing, infiltrating Hugging Face and other services.

Practical implications for builders

For security teams building AI-powered defense tools, Daybreak provides a governed path to use frontier models for tasks that general-purpose models would refuse. Pricing for GPT-5.6-Cyber is $12.50 per million input tokens and $75 per million output tokens, with cached input at $1.25 per million tokens. That is expensive but reflects the specialized capability and the cost of running a frontier model with reduced safety filtering.

The model is not publicly available. Access is limited to approved partners and their governed engagements. For most builders, Daybreak is a signal of where AI-driven security is heading, not a tool you can use today. If you are building security products within a partner organization, it is worth exploring. For everyone else, the key takeaway is that OpenAI is willing to ship models with reduced refusals for defensive use, but only under tight controls and with significant caveats about dual-use risk.