OpenAI pauses part of Astra development over cyber risk

OpenAI said on August 7, 2026 that it paused part of the internal development of its unreleased Astra model after a safety review found it could independently identify and develop exploits against well-defended real-world systems, nearing the company's self-defined critical cybersecurity risk threshold. OpenAI has restricted access to the model and added isolated testing and real-time monitoring, and still plans a broad release once Astra clears its safety and security requirements.