OpenAI has launched GPT-5.6-Cyber, a specialized model designed for vulnerability research and exploit development, accessible exclusively through the company’s vetted Daybreak cybersecurity program for authorized industry partners.

Key Points

• GPT-5.6-Cyber features a 95% completion rate for high-risk cyber tasks, significantly higher than the 1.5% rate of standard models with active safeguards.
• The model successfully identified two previously unknown vulnerabilities in Google’s V8 JavaScript engine, resulting in the issuance of CVE-2026-15903.
• OpenAI’s Daybreak program is now divided into "Blue" for general-purpose defensive work and "Red" for specialized offensive security testing.
• Access is strictly controlled through identity verification, legal attestations, and a mandatory requirement for hardware security keys starting September 1, 2026.
• Early commercial partners include major firms such as Cisco, IBM, CrowdStrike, Palo Alto Networks, and Accenture.

Why it Matters

This release signals a strategic shift toward providing security professionals with powerful offensive AI tools to preemptively identify and patch vulnerabilities before malicious actors can exploit them. By formalizing this access through the Daybreak program, OpenAI is attempting to balance the risks of dual-use technology with the urgent industry demand for advanced defensive capabilities.