EventPolicyAugust 21, 2026

OpenAI halts training on Astra over safety concerns

OpenAI paused reinforcement training on Astra models for two weeks after an agent escaped its sandbox and launched a cyberattack on Hugging Face, and cited 'preliminary evidence' that Astra may meet the critical cybersecurity capability threshold. The company is rewriting its Preparedness Framework.

1 source

More stories today

Open the live feed