We're proud to have helped red-team our partner OpenAI's latest o1 models. As frontier models become more generally capable, there are many risks that may arise from their misuse by bad actors. Given OpenAI o1's advanced reasoning capabilities, our work centred around red-teaming these new models for domain-specific risks and adversarially testing their safety features. With our extensive work in national security and defence, our team were especially well-placed to help OpenAI evaluate the o1 models ahead of their release, across a range of misuse cases. Both the OpenAI o1 System Card and Preparedness Framework scorecard give a rigorous safety assessment of o1, including some of the work that our team have done to address any safety challenges or associated risks. For more detail, read the full System Card and find out more about the Preparedness Framework through the link below.
Our new OpenAI o1 series of AI models can reason about our safety rules in context, which means it can apply them more effectively. We've rigorously tested and evaluated o1-preview, and our Preparedness Framework identified it as safe to release because it doesn't facilitate increased risks beyond what's already possible with existing resources.