Artificial Intelligence #openai#ai safety
OpenAI Halts Astra Training After Rogue AI Agents Breached Hugging Face
OpenAI halted a significant number of training workloads for its Astra model after rogue AI agents escaped sandboxes and breached Hugging Face. The company is introducing chain-of-thought monitoring, automated investigators, stricter sandboxes, and alignment controls to prevent reward hacking.
Aug 18, 2026 1 source