OpenAI has expanded its "Daybreak" cybersecurity program into two access tiers and released GPT-5.6-Cyber, a purpose-trained model with sharply reduced refusals for offensive security tasks. Daybreak Blue gives defenders a guardrail-tuned version of GPT-5.6 Sol for incident detection and vulnerability management, while Daybreak Red grants trusted partners — including Accenture, Cisco, Cloudflare, CrowdStrike, IBM and Palo Alto Networks — access to GPT-5.6-Cyber for authorized vulnerability research and exploit development.
The numbers show why OpenAI built a separate tier: GPT-5.6-Cyber completed 95.0% of exploit-related requests in testing, versus just 1.5% for the standard GPT-5.6 Sol model, which is tuned to refuse them. Using the new model, researchers found real, previously unknown flaws, including CVE-2026-15903, a high-severity (CVSS 8.8) out-of-bounds read/write bug in Chrome's V8 JavaScript engine that could allow arbitrary code execution via crafted HTML. Google has already patched the flaw. The same effort turned up mobile-OS privilege-escalation chains, three critical database vulnerabilities with remote-code-execution paths, and more than 400 kernel-level privilege-escalation flaws.
The release underscores the dual-use tension at the center of frontier AI: the same reasoning that finds zero-days for defenders could just as easily arm attackers. OpenAI is betting that gating the more capable model behind vetted partners and coordinated disclosure lets defenders benefit first.