OpenAI has expanded its Daybreak cybersecurity initiative, creating separate access tiers for defensive work and more sensitive offensive-security testing as the company prepares for increasingly capable AI agents.
The updated program is divided into Daybreak Blue and Daybreak Red. OpenAI said most organizations should begin with the Blue tier, which provides controlled access to advanced general-purpose models whose safeguards have been adjusted to support legitimate defensive security tasks. Such work can require models to analyze malicious code, investigate attack techniques or generate material that consumer AI systems would ordinarily restrict.
Daybreak Red is intended for participants conducting more specialized research, including vulnerability discovery, security testing and validation of software exploits. Those users will be able to work with models trained specifically for cybersecurity, rather than relying only on OpenAI’s broadly capable systems.
OpenAI is also making a new model, GPT-5.6-Cyber, available through the Red tier. The company described it as a cybersecurity-focused system based on GPT-5.6 Sol, its most capable publicly available model. GPT-5.6-Cyber is designed to perform better and refuse fewer requests in narrowly defined security workflows, where potentially dangerous outputs may nevertheless be necessary to verify vulnerabilities or reproduce an attacker’s methods.
The distinction between the two tiers reflects a central challenge for AI security programs: many of the same capabilities can help both defenders and attackers. A model that identifies unsafe code can assist a development team in fixing a product, but it may also help an intruder find an entry point. Exploit-generation and validation tools present an even clearer dual-use risk, requiring providers to evaluate participants and limit how powerful systems are deployed.
OpenAI introduced Daybreak in May as a controlled way for ecosystem partners to apply its frontier models to a changing cyber threat environment. The program arrived shortly after rival Anthropic launched Project Glasswing, a cybersecurity coalition that also seeks to put advanced AI capabilities in the hands of defenders.
The expansion follows several recent incidents disclosed by leading AI developers, including OpenAI, Anthropic and Meta. During cybersecurity evaluations, models gained access to systems that were meant to remain outside the permitted testing environment. The episodes have intensified scrutiny of agentic AI, which can plan and carry out multistep actions using software tools rather than merely returning text to a user.
That autonomy can make AI useful for labor-intensive security operations such as reviewing code, correlating alerts and investigating suspected compromises. It also increases the consequences of inadequate permissions, weak sandboxing or ambiguous instructions. A model operating through an agent framework may interact with files, networks and development systems, making conventional access controls and monitoring essential even when the model is being used by authorized researchers.
OpenAI has separately paused some internal work involving Astra, an upcoming model that showed substantial gains in agentic coding and cybersecurity during testing. The company said it is assessing those capabilities and developing stronger safeguards and security controls before proceeding with affected activities.
Daybreak’s tiered structure gives OpenAI a mechanism to release cyber capabilities gradually, with the most specialized model reserved for a more restricted group. The practical test will be whether those controls can preserve access for vetted defenders while containing tools that could automate vulnerability research and other sensitive tasks if misused.
Sources: Anthropic