OpenAI Flags Critical Cyber Risks in Astra Model

OpenAI announced that its upcoming model, Astra, is demonstrating significant advancements in agentic coding and cybersecurity. Internal evaluations suggest the model may possess capabilities that warrant classification under the 'Critical' tier of OpenAI's Preparedness Framework.

The company's announcement on August 7, 2026, highlighted that Astra's performance in identifying and developing zero-day exploits or devising novel cyberattack strategies without human intervention could meet the framework's stringent 'Critical' criteria. This threshold is defined as the ability to find and develop functional zero-day exploits for all severity levels in hardened real-world critical systems, or to devise and execute novel, end-to-end cyberattack strategies against such targets based on high-level goals.

Preparedness Framework in Focus

OpenAI first introduced its Preparedness Framework in December 2023, designed to guide the company's response to emerging AI capabilities, particularly in areas like biology, chemistry, cybersecurity, and AI self-improvement. Previous models, including GPT‑5.6‑Sol, were evaluated and categorized at the 'High' cybersecurity threshold. Astra's preliminary results, however, indicate a potential leap to 'Critical' capabilities, a development OpenAI is sharing transparently with the public and the broader safety and security communities.

StartupHub.ai data indicates OpenAI holds a strong score of 84/100, reflecting its leading position in AI development. Astra, as a specific model under development, scores 49/100 in comparison to other advanced AI projects like Isar Aerospace (60/100) and Rocket Lab (75/100), suggesting significant potential but also highlighting the need for rigorous validation, especially in sensitive domains like cybersecurity.

Enhanced Security Measures Implemented

In response to Astra's potential critical cyber capabilities, OpenAI has initiated a series of heightened security protocols. These include scaling up robustness testing for safeguards and security controls, implementing stricter access controls for higher-capability models, utilizing isolated testing environments, and enhancing monitoring for risky actions across all agentic applications of Astra. Internal activities not meeting these strengthened controls have been paused.

The company is also deploying universal monitoring for risky actions and misalignment in Astra's agentic applications. This system analyzes the model's Chain of Thought to trigger security responses and interrupt high-risk activities. This proactive approach mirrors the steps taken in June 2025 when models approached the 'High' capability threshold for biology, demonstrating a consistent application of the Preparedness Framework.

Collaboration and Future Outlook

OpenAI plans to collaborate with relevant government agencies and select AI safety organizations to conduct further testing of Astra's cybersecurity capabilities. They will also provide recommended security controls to third-party testing partners to ensure safe evaluations. The company maintains that advanced cyber-capable models should ultimately serve to bolster defenses by helping identify and address vulnerabilities before malicious actors can exploit them. This development underscores the accelerating pace of AI advancement and the critical need for robust safety frameworks and collaborative oversight in frontier AI research.