OpenAI Pauses Astra, Warns Next Model May Be Too Dangerous

OpenAI has paused development of its Astra research model after internal testing and reviews raised safety concerns, the company announced in a public statement.

The decision followed rounds of stress testing and safety evaluations that identified risks if the model were scaled further. OpenAI said the assessment found potential harms that current safeguards might not control.

The company listed several categories of concern: misuse for large-scale disinformation, use in automated cyberattacks, challenges in aligning highly capable systems with human values, and the difficulty of predicting new behaviors as models grow more complex.

OpenAI did not provide technical specifications for Astra, a timeline for resuming work, or say whether development resources would be redirected to other projects.

Employees and outside safety researchers have in recent months called for stronger guardrails around advanced AI. The announcement followed internal discussions about how to balance faster capability development with reducing the risk of serious harm.

OpenAI said it will continue safety research and work on tools to monitor and control powerful systems. Engineers and researchers will focus on measures aimed at improving model robustness and oversight while Astra’s development is paused.

“After careful review, we have decided to pause Astra’s development because the next model could be too dangerous to build,” the company wrote in its statement.

The company added it will coordinate with external experts and share more information about the pause and its safety work as appropriate. OpenAI did not say whether the pause will affect commercial products or current customers.