OpenAI Launches GPT-5.6-Cyber as Daybreak Program Expands

OpenAI is expanding its Daybreak cybersecurity program as AI models become more capable of finding and exploiting software vulnerabilities.

The company is splitting Daybreak into two access tiers, Blue and Red, with different models and safeguards for approved cybersecurity professionals.

Daybreak Blue expands access to frontier models

Daybreak Blue gives approved security professionals access to OpenAI’s frontier general-purpose models, including GPT-5.6 Sol.

OpenAI adjusts safeguards under the program to support more legitimate cybersecurity work than its standard models normally allow. Researchers can use the models for vulnerability discovery, secure code reviews, malware analysis, incident response, and patch validation.

OpenAI expects the Blue tier to cover the needs of most developers and security researchers.

Daybreak Red adds GPT-5.6-Cyber

Daybreak Red targets more advanced cybersecurity research and provides access to specialized models with stronger capabilities.

The new GPT-5.6-Cyber model will be available through this tier. OpenAI based the model on GPT-5.6 Sol but designed it to handle more dual-use cybersecurity requests that standard models would normally restrict.

Approved researchers can use GPT-5.6-Cyber for advanced vulnerability research and exploit validation. OpenAI limits access more heavily because of the model’s stronger cyber capabilities.

GPT-5.6-Cyber completes 95% of advanced requests

OpenAI’s evaluation shows a significant performance gap between GPT-5.6-Cyber and its general-purpose models.

GPT-5.6-Cyber completed 95% of advanced cybersecurity requests, compared with 57.3% for GPT-5.5-Cyber.

Standard GPT-5.6 Sol completed only 1.5% of the requests under its normal safeguards, while GPT-5.6 Sol through Daybreak Blue reached 2%.

GPT-5.6-Cyber finds real-world vulnerabilities

OpenAI says GPT-5.6-Cyber has already helped researchers identify real vulnerabilities.

The model discovered two vulnerabilities in Chrome’s V8 engine that researchers could chain together to bypass the browser sandbox. Google has since patched the issue.

GPT-5.6-Cyber has also identified hundreds of additional vulnerabilities across operating systems, databases, and other software.

The announcement comes as researchers increasingly test what autonomous AI systems can do outside controlled environments. Recently, Meta AI breached an organization during testing, while Anthropic and OpenAI models targeted real people in separate cyber evaluations.

Microsoft has also published guidelines for containing AI agents as concerns grow about agents escaping or exceeding sandbox restrictions.