OpenAI: We'll hit pause on model reinforcement learning for safety

OpenAI said it is pausing some of its reinforcement learning training for its frontier models to meet alignment, security and monitoring standards.

In a blog post, OpenAI outlined the changes to its safety practices following a well-publicized incident where its AI models breached Hugging Face.

• UK's AISI finds 19 instances where Anthropic's Mythos, OpenAI's GPT-5.6 Sol tried attacks
• Anthropic said Claude hacked three companies: Real worry or marketing?
• Hugging Face OpenAI attack postmortem points to lack of AI agent visibility
• OpenAI details how its models went rogue, attacked Hugging Face
• Hugging Face defends agentic AI attack with Z.ai's GLM 5.2

The latest installment of OpenAI's safety changes landed following separate posts on ChatGPT for teens and security practices. OpenAI previously noted that its Astra model may have critical cyber capabilities.

OpenAI said:

"As models become more capable, the risks associated with developing and testing them internally also grow. Our standards for monitoring, alignment, and security must stay ahead of those risks. We wanted to take the time necessary to meet those standards, so we temporarily slowed the pace of scaling. This included a two-week pause in reinforcement learning (RL) training on our latest models intended for deployment while we further hardened and red-teamed our research environments and expanded the coverage of our monitoring systems. Our largest planned frontier RL run remains on hold while we conduct smaller-scale training and evaluations to assess model behavior, validate our safeguards, and establish more evidence of alignment before proceeding."

The extra time will be used to ensure AI systems behave as intended and under human oversight.

OpenAI said its development for more capable models, the ones that will likely be used for cybersecurity defenses, will have beefed up safeguards for monitoring, alignment and security measures.

The company outlined the following changes to its safety practices:

• Stronger isolation and sandboxes for workloads that execute model-generated or untrusted code.
• Network isolation to prevent models from the internet.
• Continuous security testing includes reducing privileges, ensuring trust boundaries and improved log collection and monitoring.
• An expansion of chain-of-thought monitoring for long running sessions and training runs. OpenAI added a multistage monitoring system that runs at every sampled token and escalates concerns.