OpenAI Flags Potential Critical Cybersecurity Risks in Upcoming Astra Model, Tightens Safety Controls

San Francisco, 8 August 2026 – Artificial intelligence research laboratory OpenAI has announced a temporary pause on certain internal development activities involving its next-generation model, Astra, after preliminary evaluations revealed advanced cybersecurity capabilities that could meet the threshold for critical risk. Under the company’s Preparedness Framework, a model is classified at the critical capability level if it demonstrates the ability to autonomously identify, develop, and execute zero-day exploits against hardened software systems without human intervention.

In an official security disclosure released on Friday, OpenAI stated that recent evaluations showed significant advancements in agentic coding and autonomous problem-solving within Astra. While previous flagship models, including GPT-5.6-Sol, were categorized at the high risk threshold, preliminary benchmarking on Astra indicated performance levels that prevented safety teams from ruling out critical cyber capabilities. The disclosure follows broader industry discussions at the Black Hat security conference regarding the rapid evolution of autonomous AI agents capable of executing offensive cyber operations.