OpenAI is widening its Daybreak cybersecurity program with a sharper split between defensive work and offensive security research. The centerpiece is GPT-5.6-Cyber, a specialized model built for Daybreak Red users who are trying to find vulnerabilities and validate exploits before attackers can use AI-powered tools at scale.
The move reflects a clear concern from OpenAI: threat actors are expected to use AI more often in cyberattacks, including fully autonomous ones. Daybreak is meant to give defenders and trusted researchers earlier access to stronger tools, while adding verification, monitoring, and security requirements around that access.
How Daybreak is changing
The expanded Daybreak program now has two access tiers. Daybreak Blue is designed for authorized defensive work, while Daybreak Red is aimed at security researchers working closer to exploit development and penetration testing.
Daybreak Blue gives users access to GPT-5.6 Sol with safeguards adjusted for defense-focused cybersecurity tasks. The source describes those tasks as vulnerability detection, malware analysis, and incident response.
Daybreak Red goes further. It is built for vulnerability research, exploit validation, and penetration testing. Through this tier, users can access GPT-5.6-Cyber, which is based on GPT-5.6 Sol but trained specifically for offensive security purposes.
That distinction matters because many general-purpose AI systems refuse to answer sensitive security prompts. GPT-5.6-Cyber is intended to operate in a narrower, controlled setting where advanced security work is expected rather than blocked by default.
What GPT-5.6-Cyber is built to do
GPT-5.6-Cyber was trained to perform better on tasks such as finding zero-day vulnerabilities and building exploit chains. In practical terms, that means the model is meant to help researchers move from identifying a weakness to understanding whether it can be turned into a working exploit.
According to OpenAI, the model rarely refuses security-related queries that other models block by default. In an internal benchmark called "Advanced Cybersecurity Completion Rate," GPT-5.6-Cyber answers 95 percent of queries covering scenarios such as exploit chain development, authentication bypass, and privilege escalation.
The comparison in the source is stark. GPT-5.6 Sol with safety measures turned on reaches 1.5 percent on that benchmark. Daybreak Blue reaches 2 percent. GPT-5.5-Cyber reaches 57.3 percent.
OpenAI also tested the models on a WebSocket authentication bypass for an internal admin panel. In that case, only GPT-5.6-Cyber on Daybreak Red produced working exploit code. Every other variant refused to respond.
On ExploitGym, a benchmark for measuring how well models turn known vulnerabilities into working exploits, GPT-5.6-Cyber also outperformed both GPT-5.6 Sol and GPT-5.5-Cyber.
Why access controls are central
Because GPT-5.6-Cyber is designed to answer requests that other models usually block, OpenAI is pairing the new capability with account and workflow controls. Entry into either Daybreak tier requires identity verification, account security measures, monitoring, and legal declarations.
Hardware security keys become mandatory for all Daybreak accounts on September 1, 2026. OpenAI also recommends that security workflows run in isolated sandbox environments.
The company further recommends using Auto-Review mode in Codex. The source describes this as a mode that checks actions requiring elevated privileges before they run.
Those requirements underline the basic tension around AI for cybersecurity. The same skills that help defenders identify and validate vulnerabilities can also be sensitive when used outside authorized research. Daybreak Red appears to be OpenAI’s attempt to make advanced exploit research possible while narrowing who can use the model and under what conditions.
Real-world vulnerability research
OpenAI has already used GPT-5.6-Cyber in real-world vulnerability research. The company says the model analyzed V8, Chrome's JavaScript engine, and found two previously unknown vulnerabilities.
Those flaws can be chained together to corrupt memory and bypass the V8 heap sandbox. Google fixed the issues after coordinated disclosure and assigned them the CVE-2026-15903 designation.
The model also reportedly found at least five vulnerabilities in a "popular mobile operating system." One of them is described as a chain of flaws that would allow an app to escalate its normally restricted access rights to full administrator privileges, giving it control of the device.
OpenAI is working with Daybreak partners and the open-source community to disclose and fix those issues.
What the rating says about the direction of AI security
Under OpenAI's Preparedness Framework, GPT-5.6-Cyber has been rated "High" for cybersecurity capabilities. It does not reach the "Critical" threshold.
The source also notes that the recently announced Astra model is "potentially" expected to reach that Critical level. That context makes GPT-5.6-Cyber important for more than its current benchmark results. It shows how quickly specialized AI systems are moving into security work that previously required rare human expertise.
For defenders, the practical implication is that vulnerability discovery and exploit validation may become faster and more automated. For security programs, the challenge is to use that speed responsibly: find weaknesses earlier, validate risks more clearly, and fix issues before attackers can put similar capabilities to work.