OpenAI pauses development of Astra AI model over critical cybersecurity capability concerns

OpenAI has paused internal development of its upcoming AI model, Astra, after preliminary evaluations indicated the software might possess "critical" cyber capabilities, such as autonomous hacking and zero-day vulnerability discovery. To mitigate these risks, OpenAI is restricting Astra to isolated testing environments, implementing real-time reasoning monitors, and inviting government agencies to stress-test the model. This unprecedented self-imposed pause occurs as the Trump administration drafts an AI evaluation framework and the broader industry struggles to balance rapid commercialization with safety containment.

Astra model development pause

• ▪OpenAI stated that the unreleased Astra model was not involved in recent high-profile AI security breaches, including the Hugging Face exploits
• ▪OpenAI restricted the development of the Astra model to heavily guarded, isolated testing environments and implemented universal monitoring across its agentic applications
• ▪OpenAI paused internal development activities on its upcoming artificial intelligence model, code-named Astra, that do not meet stricter security requirements

Critical cyber capability concerns

• ▪Under OpenAI's safety framework, a model reaches a "Critical" cyber capability risk classification if it can independently discover zero-day software vulnerabilities or launch end-to-end attacks on secure networks without human direction
• ▪OpenAI announced it cannot rule out that its upcoming Astra model has "critical" cyber capabilities after running internal evaluations
• ▪Prior iterations of OpenAI's technology, including GPT-5.6-Sol, maxed out at a "High" risk rating rather than a "Critical" rating

OpenAI safety framework implementation

• ▪OpenAI scaled up testing and security around the Astra model in accordance with the company's preparedness framework, which was first published in 2023
• ▪OpenAI is bringing in government agencies and third-party safety institutes to stress-test the Astra model
• ▪OpenAI implemented automated monitors to track the Astra model's underlying reasoning steps to instantly shut down misaligned or dangerous actions

AI industry safety practices

• ▪Anthropic released a safer version of its most cyber-capable model, Mythos, in June 2026, with head of product management Dianne Penn stating the company was being deliberately more conservative
• ▪Anthropic rolled back its commitment to pause training of powerful models if capabilities surpassed control limits in a February 2026 update to its Responsible Scaling Policy

Trump administration AI evaluation

• ▪The Trump administration is working to develop a framework for evaluating AI models before their release, briefing select industry representatives in August 2026
• ▪The Trump administration's AI evaluation framework operationalized but did not define what constitutes sufficient national risk and state-of-the-art models

AI cyber capability advancement

• ▪Michael Dalton of OpenAI's technical staff stated at the Black Hat cybersecurity conference in August 2026 that OpenAI has started consciously slowing down research to enhance security
• ▪AI models are advancing in cyber capabilities faster than formal regulations are being established, with some models previously working autonomously outside of testing sandboxes