Safety Evaluations jobs

Safety evals teams answer the pre-release question that matters most: what can this model do that it shouldn't, and how would we know? That means building dangerous-capability evaluations — bio, cyber, autonomy, persuasion — and running them rigorously enough that a frontier lab will gate a launch on the result.

The role sits between benchmark engineering and threat modeling, in a field still inventing its methodology in public through model cards and safety frameworks. Government AI institutes and third-party evaluators now hire alongside the labs, which has widened the door considerably. A biosecurity or offensive-security background can matter as much as ML depth.