OpenAI has officially released GPT-5.6-Cyber, marking its first foray into what it describes as an "offense-grade" cybersecurity AI. Launched on August 10, 2026, the model is built upon the GPT-5.6 Sol reasoning architecture but has been specifically fine-tuned for discovering zero-day vulnerabilities and constructing exploit chains. While the standard Sol model only answers 1.5% of advanced cybersecurity prompts in internal benchmarks, GPT-5.6-Cyber achieves a 95% completion rate. Its effectiveness was demonstrated by identifying two previously unknown vulnerabilities in Google’s V8 JavaScript engine, now tracked as CVE-2026-15903. To mitigate risks, OpenAI has restricted access to a vetted group of researchers through its "Daybreak Red" program, which requires rigorous identity verification and continuous monitoring. While the model is currently rated as "High" risk under OpenAI's Preparedness Framework, its release signals a shift in the dual-use nature of AI, where offensive capabilities are provided to authorized defenders to maintain parity with potential bad actors.