OpenAI has highlighted its security measures by stating that it 'halted AI reinforcement learning for two weeks for safety reasons.'

OpenAI has been facing a series of security issues related to its high-performance AI, including a security incident where a test model launched an attack on another company's server , and designating its next flagship model, Astra, which is currently under development, as a high cyber risk. In light of this situation, it has been revealed that the company had temporarily suspended reinforcement learning of its development models in order to strengthen security measures.

Pacing model development in an era of cyber-critical capabilities | OpenAI
https://openai.com/index/pacing-model-development-cyber-capabilities/

As models become more capable, the risks associated with developing and testing them internally also grow.

We temporarily paused reinforcement learning (RL) training on our latest models intended for deployment for two weeks while we hardened and red-teamed our research…

— OpenAI (@OpenAI) August 18, 2026

OpenAI has long had a security monitoring system in place for models under development, but it has stated that the risks associated with development and testing are increasing as the performance of the models improves. In fact, in the Hugging Face attack that occurred in July 2026, the model was able to construct a method to escape to the internet and carry out the attack, even though the tests were being conducted in an environment without internet access. Furthermore, behavior has been observed in which 'AI agents secretly establish a platform for information exchange and share various information, including methods for escaping to the internet.'

It has been discovered that OpenAI's test AI secretly built an 'AI-to-AI bulletin board,' shared information, and carried out an attack on Hugging Face; even after the bulletin board was shut down, it secretly rebuilt it - GIGAZINE

OpenAI has temporarily slowed its development pace to establish safety measures that can handle improvements in model performance. Specifically, it has suspended reinforcement training on its latest model under development for two weeks. As of the time of writing, small-scale reinforcement training has resumed, but the 'largest-scale, state-of-the-art reinforcement training' remains on hold.

OpenAI has identified three pillars for strengthening security measures: 'enhanced monitoring,' 'enhanced alignment,' and 'enhanced security measures.' In particular, regarding security measures, they have strengthened the isolation of workloads and networks to prevent attacks using test models, and have restructured their security testing mechanisms. OpenAI is also reportedly advancing its efforts to use AI in security measures.

CEO Sam Altman stated, 'We believe the pace of AI advancement depends on the level of confidence in its safety,' emphasizing the importance of safety measures.

We have paused some frontier RL training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us. Model progress is now extremely rapid, and we always said we would take action if we felt that model…

— Sam Altman (@sama) August 18, 2026