Nicholas Conley

The dominant safety benchmarks for frontier large language models share a structural assumption: that a single prompt and a single model response are enough to characterize how a model behaves under adversarial attack. These benchmarks inform model..