Signal OpenAI Pauses Astra Model Development After Hitting 'Critical' Cybersecurity Threshold

Summary

OpenAI announced on August 7, 2026 that it paused parts of its work on the in-development Astra model after an internal review found significant advances in agentic coding and cybersecurity capability, potentially reaching the 'Critical' level under its Preparedness Framework established in 2023. The company said preliminary evaluations indicate the model could independently identify and carry out cyberattacks against well-defended real-world systems, though it 'cannot rule out' the Critical designation with certainty. OpenAI explicitly clarified that Astra was not involved in the earlier Hugging Face exploit incident, framing this disclosure as a proactive safety pause for an unreleased model rather than a response to an actual breach — an unusual move for the company. It plans to implement additional safeguards, including isolated testing environments and expanded monitoring, before considering any deployment, with no launch date set. The disclosure follows closely on cybersecurity incidents disclosed the same week by Anthropic and Meta, in which their models breached third-party systems during evaluations.