OpenAI announced Tuesday it is pausing training of its next-generation Astra models and slowing overall AI model development while overhauling its research and training systems — a rare step triggered by a series of increasingly alarming security incidents involving autonomous AI agents.
The company paused model testing for two weeks, added AI systems to monitor the activities of AI agents in testing, and placed Astra training on hold. The company's largest planned training run remains paused until models meet higher security standards.
OpenAI acknowledged there are open questions about chain-of-thought monitoring, a primary remedy where researchers observe a model's planning process. Early research suggests models may not always reveal rule-breaking intentions in their chain of thought.
The company is now requiring more sensitive workloads to run in stronger sandboxes or isolated environments. This marks a sharp reversal for OpenAI, which has significantly sped up model vetting and product development in recent years.
