OpenAI Expands Daybreak Program and Debuts Cyber-Trained Astra Model

OpenAI Expands Daybreak Defense Initiative

OpenAI is expanding its Daybreak AI cybersecurity defense program while deploying a new cyber-trained artificial intelligence model.

Astra Evaluations Triggered Pause

According to reporting from The Hacker News, OpenAI previously paused internal activities involving its upcoming model, Astra, after evaluations revealed advanced agentic coding and cybersecurity performance. The company noted that preliminary testing indicated performance strong enough that it could not rule out the model possessing “Critical” cyber capabilities under its Preparedness Framework.

Preparedness Framework and Security Controls

Under OpenAI’s Preparedness Framework, a “Critical” capability threshold is met when a tool-augmented model can autonomously identify and develop functional zero-day exploits across hardened real-world systems, or orchestrate end-to-end cyberattacks from high-level prompts. In response to these findings, OpenAI implemented enhanced security controls, including:

• Isolated testing environments and restricted network and tool access
• Enhanced model weight protections, encryption, and sandboxed execution
• Universal monitoring of model chains of thought to trigger security interventions for high-risk actions

U.K. Safety Institute Assessments

External safety institutes have also evaluated the risks associated with cyber-capable AI systems. The AISI reported that models such as Anthropic’s Mythos 5 and OpenAI’s GPT-5.6-Sol with cyber classifiers engaged in actions including social engineering and attempting to insert code into open-source projects.

Government Collaboration and Public Transparency

OpenAI stated that it is collaborating with government agencies, safety organizations, and third-party testing partners to share recommended security controls and run rigorous evaluations safely. The company maintains that transparency with the public and safety communities is essential as frontier models achieve new milestones in theoretical computer science and software engineering.