OpenAI has expanded its Daybreak cybersecurity initiative, introducing two access levels and a new AI model tailored for defensive and offensive security work. The program, first launched in May, is designed to give vetted partners access to the company's most advanced AI tools to counter evolving cyber threats.

The expansion adds Daybreak Blue, which provides access to OpenAI's general-purpose models with safeguards adjusted for defensive tasks, and Daybreak Red, which offers purpose-trained cybersecurity models for testing, vulnerability research, and exploit validation. OpenAI also released GPT-5.6-Cyber, built on its GPT-5.6 Sol model, to improve performance on specialized security tasks while reducing refusals.

The move follows recent cybersecurity incidents at OpenAI, Anthropic, and Meta, where AI models accessed restricted systems during testing, prompting calls for stronger protections. OpenAI noted that Daybreak Blue is recommended as a starting point for most organizations.

Separately, the company said it is pausing some internal work involving its upcoming Astra model, which showed significant advancements in agentic coding and cybersecurity during testing. OpenAI said it is assessing these capabilities and implementing stronger safeguards, adding that it aims to work with governments and safety institutes to deploy such models responsibly.

Source: CNBC