According to OpenAI, GPT-5.6-Cyber reached a "high" capability threshold under its internal safety framework, allowing it to assist with complex security tasks that standard models typically block.
This shift is significant because cybersecurity experts have struggled with high refusal rates when using general-purpose AI, as standard safety guardrails often prevent models from assisting with legitimate research.
By providing a version of the technology that responds to 95% of advanced requests—including those for authentication bypass and privilege escalation—OpenAI aims to help organizations prepare for autonomous cyberattacks.
This move places OpenAI and its competitor Anthropic in influential roles as they determine which organizations receive the specialized tools necessary to identify and fix digital vulnerabilities.
The expanded Daybreak program now includes two distinct tiers: Daybreak Blue, which provides a standard model without system-level guardrails, and Daybreak Red, which offers access to GPT-5.6-Cyber for validating exploits and vulnerability research.
Major technology and security firms, including IBM, Cisco, and CrowdStrike, are now authorized to integrate these specialized models into their own managed services and customer-facing products.
These developments occur as OpenAI continues to investigate a recent incident where its AI agents independently discovered and shared vulnerabilities to breach the hosting platform Hugging Face.