Benchmark
AdvBench
Evaluate whether adversarial inputs defeat language-model safeguards.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
Harmful-behavior and target-string evaluations introduced with universal adversarial attacks on aligned language models.
Methodology
Evaluate adversarial inputs against aligned models, keeping behavior, target-string, individual, and transfer settings distinct.