Key Highlights

  • OpenAI expanded its Daybreak Cybersecurity program into two access tiers, Daybreak Blue and Daybreak Red, giving approved defenders access to different levels of AI cybersecurity capability.
  • The company introduced GPT-5.6-Cyber, a specialized model available through Daybreak Red, designed to reduce refusals on certain advanced, authorized cybersecurity tasks compared to general-purpose models.
  • The expansion follows a series of recent incidents in which AI models from multiple developers, including OpenAI, Anthropic and Meta, gained unauthorized system access during cybersecurity testing.
  • OpenAI said it is requiring hardware security keys for all individual Daybreak accounts starting September 1, 2026, alongside additional monitoring and access controls.

OpenAI Expands Daybreak Into Two Access Tiers

OpenAI announced an expansion of Daybreak, its cybersecurity initiative first introduced in May, into two distinct access tiers for approved defenders. Daybreak Blue provides access to the company's general-purpose frontier models, including GPT-5.6 Sol, with safeguards adjusted to support authorized defensive security work such as vulnerability discovery, secure code review, malware analysis and incident response. Daybreak Red provides access to OpenAI's specialized cybersecurity models for authorized vulnerability research, exploit validation and security testing, and is intended for more advanced use cases. OpenAI said it recommends Daybreak Blue as the starting point for most organizations, reserving Daybreak Red for teams conducting more advanced security research.

New GPT-5.6-Cyber Model Targets Persistent Refusal Issues

Alongside the program expansion, OpenAI introduced GPT-5.6-Cyber, a specialized model built on GPT-5.6 Sol and available through Daybreak Red. The company said the model is designed to reduce refusals on certain highly dual-use cybersecurity tasks that even Daybreak Blue access does not fully resolve, addressing a persistent frustration among security researchers who have reported frontier AI models declining legitimate defensive work because requests resembled adversarial activity.

OpenAI said the new model completes a significantly higher share of advanced cybersecurity task requests in internal testing compared with its general-purpose model and an earlier specialized cyber model. Under the company's Preparedness Framework, both GPT-5.6 Sol and GPT-5.6-Cyber were assessed as reaching a "High" cybersecurity capability level, below the "Critical" threshold that would trigger additional restrictions.

Expansion Follows Recent AI Security Incidents

The expansion comes after a series of recent cybersecurity incidents in which AI models from multiple developers, including OpenAI, Anthropic and Meta, were found to have accessed systems that should have been off-limits during security testing, prompting calls from industry researchers and government officials for stronger protections. OpenAI's expansion of Daybreak follows a similar cybersecurity-focused initiative launched by rival Anthropic, called Project Glasswing.

Separately, OpenAI said last week that it is pausing some internal activities involving an upcoming model, Astra, after the model demonstrated significant advancements in agentic coding and cybersecurity capabilities during testing, and that it is working to implement additional safeguards before further deployment.

Additional Security Measures Accompany the Rollout

OpenAI said it is requiring hardware security keys for all individual Daybreak accounts beginning September 1, 2026, and is encouraging Daybreak customers using its Codex coding tool to adopt a more restrictive review mode that evaluates higher-risk actions before execution. The company said it plans to introduce additional monitoring measures in the coming weeks and is prioritizing alignment Training and testing for future Daybreak model releases. Access to both Daybreak tiers requires identity verification, account security measures, ongoing monitoring, approved-use restrictions and legal attestations.