Definition
Plain language
A test set for checking whether an AI refuses harmful requests when it should.
As stated in the literature
A safety benchmark measuring single-turn refusal across categories of unsafe requests; cited as representative of evaluation that may not capture multi-turn persuasion vulnerabilities.
Why it matters: It measures refusal across many categories, though, by testing single requests, it may miss weaknesses that surface over a longer back-and-forth.
For example, it sends a model a range of harmful requests and checks whether it correctly declines each one.
Heard on the show
“There's a whole ecosystem of public benchmarks — XSTest, SORRY-Bench, others — that measure single-turn refusal.”Episode 045 — When a Frontier Model Talks Its Own Twin Into Climate Denial