OpenAI ships GPT‑5.6‑Cyber, a red‑team‑grade model for enterprise defense

OpenAI has begun offering a specialized security model, GPT‑5.6‑Cyber, to vetted enterprise and research teams, positioning it as a tool for serious defensive work rather than a general‑purpose assistant. Distributed through the company’s Daybreak Red program, the model is designed for tasks like penetration‑testing support, exploit‑chain design, and privilege‑escalation analysis in controlled environments. According to summaries shared with customers, GPT‑5.6‑Cyber can solve roughly 95% of advanced exploit‑chain and escalation challenges used in internal evaluations, and has already helped uncover two previously unknown zero‑day vulnerabilities in Chrome’s V8 JavaScript engine.

That level of capability highlights a long‑running dilemma: you can’t build a model that’s genuinely useful for hard‑core defenders without also creating something that attackers would love to get their hands on. OpenAI says it is mitigating the risk through strict customer vetting, detailed logging, and in‑model safety constraints, but critics argue that powerful offensive knowledge inevitably leaks over time. As governments move toward tighter oversight of “high‑risk AI,” especially in cyber operations, GPT‑5.6‑Cyber is likely to become a test case for how frontier‑grade hacking assistance can be governed without freezing legitimate security research.