OpenAI gives vetted security professionals an AI that answers where others refuse

OpenAI has expanded its Daybreak cybersecurity program and introduced a new model, GPT-5.6-Cyber. The model is intended only for vetted security professionals who need to find and fix weaknesses in IT systems. The information here is based on reporting from CSO Online and The Decoder, not on OpenAI's own announcement.

Two access tiers

Daybreak now has two tiers. The first, Blue, gives vetted defenders access to general frontier models such as GPT-5.6 Sol for authorized defensive work. The second, Red, gives access to specialized cyber models for vulnerability research, exploit validation (code that takes advantage of a vulnerability) and security testing.

Access is limited to approved individuals and organizations. Identity verification is required, and usage is monitored. Hardware security keys are planned to become mandatory from September 1, 2026. According to The Decoder, the requirements also include legal attestations. Isolated test environments (sandboxes) are recommended, and the Codex tool has an Auto-Review mode that checks privilege escalation.

Answers more often instead of refusing

In an internal test of how often models answer advanced cyber requests instead of refusing them, GPT-5.6-Cyber answered 95 percent. GPT-5.6 Sol under Daybreak Blue answered 2 percent. According to The Decoder, its predecessor GPT-5.5-Cyber answered 57.3 percent of sensitive security requests, and the standard version of GPT-5.6 Sol answered 1.5 percent. The figures come from OpenAI's own tests and have not been independently verified.

The model is built on GPT-5.6 Sol and trained to refuse less on certain high-risk dual-use requests, meaning those that can be used for both defense and attack, when users are vetted. In OpenAI's Preparedness Framework, it is rated "High" for cyber capabilities, which places it below the "Critical" level.

Vulnerabilities found

OpenAI says the model found two previously unknown flaws in Google's V8 JavaScript engine. Together, they can allow memory corruption and an escape from V8's heap sandbox. The flaws have been reported to Google through coordinated vulnerability disclosure. According to The Decoder, one of them has been registered as CVE-2026-15903.

The Decoder also reports that the model found at least five vulnerabilities in a widely used mobile operating system, including a chain that gives unrestricted administrator access. The system is not named.

Forrester analyst Biswajeet Mahapatra says: "CISOs should assume that the time between vulnerability discovery and exploitation will continue to shrink." This means security chiefs should expect less time between the discovery of a weakness and its exploitation.