On Tuesday, OpenAI announced a raft of security measures following the Hugging Face incident, aimed at better monitoring and alignment during model development.

The new policies involve enhanced surveillance during the testing phase and greater emphasis on safety protocols post-training. While these aren't directly linked to the Hugging Face breach, they reflect growing concerns over AI capabilities.

OpenAI’s VP of research, Amelia Glaese, highlighted that controls will become stricter as models grow more complex, with larger models under heightened scrutiny. The company has also paused reinforcement learning and is conducting smaller-scale evaluations before resuming operations.

The new system includes stronger network isolation practices and a robust monitoring system to detect unauthorized behaviour within 30 minutes. However, specifics remain vague, leaving some questions unanswered about how these measures will be implemented in practice.