OpenAI Halts Advanced AI Model Training to Assess Cyber Risks

OpenAI announced a temporary slowdown of its most advanced AI model development after internal tests revealed concerning cyber capabilities. The move pauses reinforcement‑learning (RL) training for about two weeks, while smaller training runs and evaluations continue to test safety measures.

What’s being halted

The largest planned training run remains on hold, but some Astra model training resumed after new security safeguards were applied. A significant portion of operations stays suspended until they can be moved to hardened environments.

OpenAI expanded its Preparedness Framework to cover risks that may arise during training itself, not only after deployment. The company also launched GPT‑5.6 Cyber as part of the Daybreak cyber‑defense program, giving developers tools for vulnerability discovery and code review.

Background and safeguards

Early assessments of Astra showed cyber‑capabilities that could reach “critical” levels by OpenAI’s standards, prompting tighter isolation of tasks, reduction of permanent privileges, and removal of shared services that might be weak points. A new monitoring system now inspects every token generated to flag suspicious behavior before escalating to automated investigation pipelines.

This monitoring consumes roughly 20 % of the inference compute budget, increasing operational costs but strengthening protection against unauthorized access and data theft.