OpenAI is expanding Daybreak, its controlled-access cybersecurity program, with two tiers for approved defenders.
Daybreak Blue provides GPT-5.6 Sol with safeguards adjusted for authorized security work, while Daybreak Red provides specialized models for higher-risk research, including the new GPT-5.6-Cyber.
GPT-5.6-Cyber is designed to handle security requests that general-purpose models often refuse, particularly vulnerability research and exploit development. OpenAI says the goal is to give vetted defenders access to stronger capabilities before AI-enabled attacks become faster and more autonomous.
Why It Matters: Daybreak shows how powerful cyber AI could be deployed under controlled access. OpenAI is pairing fewer model restrictions with tighter requirements around who can use the technology and what agents can do. Its research also shows AI moving closer to sustained vulnerability discovery in real software.
- Daybreak Separates Everyday Defense From Higher-Risk Security Research: OpenAI recommends Daybreak Blue for most defenders. It removes safeguards that can interfere with legitimate work such as incident response, malware analysis, vulnerability management, and code review. Daybreak Red gives approved researchers access to models designed for exploit validation and advanced vulnerability research. GPT-5.6-Cyber sits in this tier because it has been trained to comply with requests that GPT-5.6 Sol may still refuse.
- GPT-5.6-Cyber Completes Far More Advanced Security Requests: On OpenAI’s internal evaluation covering exploit chains and similar advanced scenarios, GPT-5.6-Cyber completed 95% of requests. GPT-5.5-Cyber reached 57.3%, while GPT-5.6 Sol remained near 2% with or without Daybreak Blue. The specialized model also performed better on ExploitGym and OpenAI’s zero-day evaluation. GPT-5.6 Sol retained an advantage in some tests, including vulnerability report writing and the standard 300-turn ExploitBench setting, showing that GPT-5.6-Cyber is optimized for specific research workflows instead of replacing the general model across security work.
- Real-World Research Has Uncovered Serious Software Flaws: OpenAI says GPT-5.6-Cyber helped researchers discover two previously unknown vulnerabilities in V8, Chrome’s JavaScript engine, that could be chained to corrupt memory and escape the V8 heap sandbox. Google fixed the reported issue and assigned CVE-2026-15903. OpenAI also reports finding vulnerabilities in a popular mobile operating system and database, plus more than 400 privilege-escalation issues in an operating-system kernel. Work to disclose and remediate those findings is underway.
- Cyber Capabilities Remain Below OpenAI’s Critical Threshold: GPT-5.6-Cyber received a High cybersecurity capability rating under OpenAI’s Preparedness Framework, matching GPT-5.6 Sol’s classification. OpenAI says specialized training improved selected capabilities without reaching its Critical threshold. The company plans to publish a system card with further evaluations at a later date. OpenAI also clarified that GPT-5.6-Cyber was not involved in the Hugging Face security incident.
- Stronger Capabilities Come With Tighter Operational Controls: Daybreak requires approval and identity verification, with monitoring and restrictions on authorized use. Individual participants must adopt hardware security keys starting September 1, 2026. OpenAI is also pushing Codex users toward auto-review mode, which can inspect elevated actions before execution. Its guidance emphasizes isolated environments and clear permission boundaries, putting more responsibility on controls surrounding the agent when model-level restrictions are reduced.
Go Deeper -> Expanding Daybreak as the Cyber Defense Window Narrows – OpenAI
OpenAI expands Daybreak cybersecurity initiative as AI agent threats evolve – CNBC
