SaferAI
A nonprofit that independently evaluates AI models for dangerous capabilities and rates how seriously labs manage risk.
SaferAI is a nonprofit that tests AI models from the outside. Rather than taking a lab’s word for how safe its model is, SaferAI runs its own evaluations, often through the public API, and publishes what it finds. It also grades AI companies on their risk management practices, which has made it a regular and occasionally unwelcome presence in the industry’s news cycle.
Its 2026 evaluation of the open-weight model GLM-5.2 was a good example of why independent testing matters: the model refused none of the offensive cyber or biology tasks it was given, while a comparable closed model refused so consistently the test could not be finished. Executive director Henry Papadatos summarised the point as “the frontier of capability is not the frontier of risk”.