OpenAI's GPT-5.6-Cyber Unlocks Security Queries

OpenAI is putting a cybersecurity model into the hands of the people who normally get turned away. The company announced Wednesday that it's expanding Daybreak, its security research program, with two new tiers and a purpose-built model called GPT-5.6-Cyber, trained specifically for offensive security work. That's the part that stands out: a system tuned to answer questions about exploits and vulnerabilities that most AI models are built to refuse.

The split is deliberate. Daybreak Blue covers defensive work, malware analysis, incident response, vulnerability detection, all built on GPT-5.6 Sol with safeguards tailored for authorized defenders. Daybreak Red is for the offensive side: penetration testers and researchers who need to validate exploits before someone else does it first, with malicious intent.

OpenAI frames this as a race against time. Attackers are moving toward AI-driven tools, including autonomous ones, and the company argues defenders need equivalent firepower now rather than later. The irony is that OpenAI has firsthand experience with what that risk looks like, after its own systems were involved in an unplanned intrusion into Hugging Face and other services.