OpenAI announced on Monday that it is expanding its Daybreak cybersecurity initiative, designed to provide participants with the "right capabilities" to defend against attackers. The company first launched Daybreak in May, around the same time its primary competitor, Anthropic, unveiled a cybersecurity alliance called Project Glasswing, capturing significant attention from Wall Street and the U.S. government. OpenAI positioned Daybreak as a way for ecosystem partners to leverage its most advanced AI models to adapt to the rapidly evolving threat landscape.

OpenAI stated Monday that it is restructuring the program into two distinct access levels: Daybreak Blue and Daybreak Red. This expansion comes after several AI developers, including OpenAI, Anthropic, and Meta, disclosed multiple cybersecurity incidents in recent weeks. In each case, AI models accessed systems that should have been restricted during security testing, prompting industry researchers and government officials to call for stronger safeguards.

"As the threat landscape evolves, we are putting frontier intelligence into the hands of trusted defenders before attackers can deploy offensive AI at scale," OpenAI posted on X platform on Monday.

Where to start

In a blog post, OpenAI explained that Daybreak Blue will grant users special access to its advanced general-purpose models, with safety measures adjusted to allow for defensive security work. Daybreak Red participants go a step further, utilizing OpenAI's "specially trained cybersecurity model" for security testing, vulnerability research, and exploit validation. OpenAI also introduced a new AI model, GPT-5.6-Cyber, which will be available to Daybreak Red users. The company said this model is built on its most powerful publicly available product, GPT-5.6 Sol, but is specifically designed to enhance capabilities and reduce refusal responses in certain cybersecurity tasks. OpenAI suggested that for most organizations, Daybreak Blue is the better starting point.

Model capabilities and future safety

In recent months, OpenAI has repeatedly warned about the growing cyber capabilities of its AI models. Last week, the company said it was pausing certain internal activities involving its upcoming Astra model, which demonstrated "significant advances in agentic coding and cybersecurity" during testing. OpenAI stated it is working to evaluate these capabilities and implement stronger safeguards and controls. "We are committed to working with governments, security agencies, and civil society to ensure that the frontier capabilities of Astra and subsequent models are deployed responsibly and benefit humanity broadly," OpenAI said.