It seems that every day brings new reports of AI agents “going rogue.” Whether breaking into Hugging Face, attacking fitness websites, or creating fake profiles for social engineering attacks, a growing number of AI models are behaving like malicious attackers.
As a result, the labs developing these AI models are expanding their cybersecurity protection products. This week, OpenAI announced the expansion of Daybreak, a cyber defense service it launched earlier this year. Shortly before that, Anthropic also released Mythos, an AI model focused on cybersecurity.
Daybreak is a service that integrates models, tools, and workflows to provide cyber defenders with access to them. The expansion adds a new AI model specifically designed for cyber defense work.
On Monday local time, OpenAI said that Daybreak would now consist of two tiers: Blue and Red. Both tiers will allow approved customers to access frontier cybersecurity models that OpenAI has made available on a restricted basis.
“Frontier models” refers to the most advanced AI models currently available, and they have long been controversial. Previously, the Trump administration sought to work with AI companies to advance the deployment of such models, reportedly for security reasons. In the past, OpenAI has placed numerous safeguards on these models, limiting the range of operations customers can use them to perform.
Blue is the more basic of the two tiers. It provides a range of cybersecurity services, including incident response, malware analysis, and patch validation. OpenAI calls Blue “the recommended starting point for most defenders,” meaning that this tier may already be sufficient for most enterprises.
Red, by contrast, offers a broader and potentially riskier set of tools. OpenAI will provide users with “specially trained cybersecurity models” for security testing and vulnerability research.
The Red tier also includes a new model, GPT-5.6-Cyber, which is available only to Red users. OpenAI said that GPT-5.6-Cyber is built on GPT-5.6 Sol and offers enhanced capabilities for certain specialized cybersecurity tasks.
At present, GPT-5.6-Cyber is available only to “trusted customer partners,” reportedly including companies such as Accenture, IBM, CrowdStrike, and Cloudflare.
As threats from AI agents increase rapidly, critics have also pointed out that these security risks provide AI labs with marketing opportunities. OpenAI is clearly promoting its upgraded Daybreak service in this way.
“Cybersecurity is changing rapidly — threat actors will increasingly use AI to launch cyberattacks at unprecedented speed and scale, including fully autonomous attacks,” OpenAI said in a blog post. “As these capabilities proliferate, the window of time defenders have to prepare is shrinking.”
At the same time, enterprises still want to purchase protection services from the AI labs that best understand the security risks, because these companies are the first to encounter and study those risks.