OpenAI Pauses RL Training for Two Weeks Over Safety Concerns

On August 18, OpenAI announced that it had paused reinforcement learning (RL) training on its latest model intended for deployment for two weeks. During this period, OpenAI hardened its research environments, conducted red-team testing, and expanded internal monitoring coverage. Some lower-risk training has since resumed, but the largest-scale frontier model RL training originally planned remains paused.

OpenAI stated that it is observing model behavior through smaller-scale training and evaluations, validating new safety measures, and accumulating more alignment evidence before deciding whether to proceed. Sam Altman reposted, saying that if model capabilities exceed safety and alignment requirements, immediate action will be taken.

Over the past few weeks, two events prompted OpenAI to slow down training. The first was a Hugging Face security incident: during an internal cybersecurity evaluation, an OpenAI model breached its isolated environment, gained internet access, and ultimately infiltrated Hugging Face's infrastructure.

The second relates to the unreleased new model Astra. Preliminary evaluations suggest that Astra may reach the 'critical' cybersecurity capability threshold as defined in OpenAI's Preparedness Framework. Previously, GPT-5.6Sol's cybersecurity capability had been rated as 'high.' Reaching the 'critical' threshold implies the model could potentially autonomously discover zero-day vulnerabilities or autonomously design complete attack plans.

OpenAI summarized its safety system as three mutually reinforcing components: monitoring, alignment, and safety measures, and anticipates that most future safety work will be done by models. Specific measures include: hardening research environments, such as workload isolation, network isolation, and continuous security testing; expanding chain-of-thought monitoring using activation classifiers and automated investigation agents; and advancing alignment research, such as improving reward models and training models to honestly describe their actions.

The monitoring system is multi-stage, aiming to alert within 30 minutes of detecting concerning activity, and pausing activity if a false positive cannot be proven. Now, tool-calling RL training and evaluations at or above the Sol capability level must have monitoring enabled. Astra goes further: all Astra + tools reasoning activities must be monitored.

The news sparked heated discussion online. Some believe this means Astra's release is delayed again, others think it indicates something is wrong given the departure of C-level executives, and some view it as a marketing tactic.

At the end of the article, OpenAI stated that frontier model capabilities are accelerating rapidly, and the ability to understand, align, and protect these models must stay ahead. But the question remains: as models continue to grow stronger, can safety capabilities always stay ahead? OpenAI's choice is to stop the models first.