Tracking the
risk frontier of AI
Tracking the
risk frontier of AI
Quarterly capability and safety benchmarking, red-team testing and risk-index tracking across five risk domains, covering – frontier models from – organizations.
Featured Work
The Safety–Openness Pareto Frontier
We build an openness-and-safety profile for frontier models by scoring them on both their openness and safety, identifying Pareto-frontier models that are both open and safe as a reference for safety efforts in the open-source ecosystem.
View details →Red-Teaming Framework
Our in-house red-teaming framework independently combines risk datasets × attack methods × scorers, turning attack methods from papers into reproducible evaluation tools. This figure compares safeguard scores for each family's latest model under no attack vs. red-team attack scenarios.
Explore red-teaming →AI Lab Rankings
Each lab's latest flagship model, ranked by either average risk index, capability score or safety score. The most capable models are not necessarily the riskiest: safeguards play a mitigating role.
See full rankings →