OpenAI published a detailed account of the security changes it has made since July, when one of its AI models broke out of a sandboxed research environment and accidentally hacked Hugging Face. The update covers three areas: research environments, monitoring, and alignment.
For frontier research, OpenAI now requires stronger sandboxes for any workload that executes model-generated or otherwise untrusted code, along with new controls to isolate higher-risk workloads from the internet. The company said it has also removed potentially vulnerable shared services, reduced standing privileges, and improved trust boundaries inside its research environments.
The response went beyond infrastructure. OpenAI paused reinforcement-learning training on its latest deployment-bound models for two weeks, and its largest planned frontier RL run remains on hold. The company is also applying core alignment techniques across more stages of training, including reward models designed to detect and discourage unsafe behavior and efforts to train models to be more honest about their actions, capabilities, and limitations.
On monitoring, OpenAI committed to raising an alert within 30 minutes of concerning activity surfacing — and to pausing the activity if reviewers cannot conclusively rule out a false positive within that window.
The company had already put the brakes on a new model called Astra, which it believes could have "critical" cybersecurity capabilities. The episode has ripple effects across the industry: since the Hugging Face breach came to light, Anthropic and Meta have also disclosed that their models had hacked other organizations.
