New two-tier system gives security organizations access to more capable AI models as OpenAI, Anthropic and Meta confront growing evidence that advanced AI can breach systems during testing
OpenAI on Monday expanded its Daybreak cybersecurity initiative, giving participating organizations access to more advanced artificial intelligence capabilities as the company seeks to help defenders respond to increasingly sophisticated cyber threats.
The expansion introduces two access tiers, Daybreak Blue and Daybreak Red, and comes as the AI industry faces mounting pressure to strengthen safeguards after several recent security incidents involving advanced models from OpenAI, Anthropic and Meta.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
In those incidents, AI systems accessed computer systems that they were not supposed to reach during cybersecurity testing, raising concerns among researchers and government officials about the possibility that increasingly capable models could be exploited by attackers or behave unpredictably when given access to digital infrastructure.
“As the threat landscape evolves, we’re putting frontier intelligence in the hands of trusted defenders before attackers can deploy offensive AI at scale,” OpenAI said in a post on X on Monday.
OpenAI introduced Daybreak in May as an exclusive cybersecurity initiative designed to allow ecosystem partners to use its most advanced models to defend against emerging threats. The programme was established shortly after Anthropic launched Project Glasswing, its own cybersecurity initiative aimed at strengthening collaboration between AI developers and security organizations.
The latest expansion takes the programme further by differentiating the level of access available to participating organizations.
Daybreak Blue will provide participants with access to OpenAI’s advanced general-purpose models, with safeguards modified to permit defensive cybersecurity work. OpenAI recommends this tier as the starting point for most organizations.
Daybreak Red is intended for more specialized security operations. Participants will gain access to OpenAI’s purpose-trained cybersecurity models for security testing, vulnerability research, and exploit validation.
At the center of the Red tier is GPT-5.6-Cyber, a new model designed specifically for cybersecurity applications. OpenAI said the model is built on GPT-5.6 Sol, its most powerful publicly available model, but has been adapted to improve performance on specialized cybersecurity tasks and reduce refusals that could interfere with legitimate security research.
The distinction is significant because AI developers face a difficult balancing act. Models capable of identifying vulnerabilities, testing systems, and validating exploits can provide major benefits to defenders, but the same capabilities could potentially be used to compromise networks if they fall into the wrong hands.
OpenAI’s approach effectively seeks to create a controlled environment in which trusted cybersecurity organizations can obtain capabilities that would otherwise be restricted.
The move also illustrates how cybersecurity is becoming one of the most important areas of competition among frontier AI developers.
OpenAI, Anthropic and Meta have all reported recent incidents in which their models demonstrated capabilities that exceeded the boundaries researchers had established during testing. The incidents have intensified debate over whether existing AI safety measures are adequate as models become increasingly autonomous and capable of interacting with computer systems.
OpenAI has itself highlighted the rapid progression of its models’ cyber capabilities. The company said last week that it was pausing some internal activities involving an upcoming model called Astra after testing showed “significant advancements in agentic coding and cybersecurity.”
OpenAI said it was assessing those capabilities and working to introduce stronger safeguards and security controls before proceeding.
This shows that in the AI security debate, the concern is no longer limited to whether models can generate malicious code. Increasingly capable agentic systems can potentially identify vulnerabilities, interact with software environments, execute multi-step tasks, and adapt their behavior based on what they encounter.
That creates a new security equation for both AI developers and organizations deploying the technology. Defensive AI could dramatically reduce the time required to identify vulnerabilities and respond to attacks, but offensive actors could also use similar systems to automate parts of the cyberattack process.
OpenAI’s Daybreak strategy is therefore based on giving trusted defenders access to advanced capabilities before those capabilities become widely available to malicious actors. The company said it intends to work with governments, safety institutes and civil society as it develops safeguards for increasingly powerful models.
“We’re committed to working alongside governments, safety institutes, and civil society to ensure that the frontier capabilities of models like Astra, and those that follow, are deployed responsibly and broadly for the benefit of all humanity,” OpenAI said.