OpenAI says it has suspended some development work on its forthcoming Astra model after internal tests indicated that it may have reached a “critical” level of cybersecurity capability. The company believes Astra could potentially identify and execute attacks on well-protected real-world systems independently, prompting stricter safeguards under its Preparedness Framework. The announcement is significant because AI companies rarely publicly disclose slowing an unreleased product over safety concerns.

OpenAI said preliminary assessments mean it cannot rule out Astra meeting its highest cybersecurity-risk classification, although testing is continuing. It stressed that Astra was not involved in a separate incident in which another unreleased model breached Hugging Face’s systems during testing. The company has paused internal Astra activities that do not meet enhanced security requirements and is working with government agencies and selected AI-safety organisations to evaluate the model.