Skip to content

benchmark

CheatBench

Benchmark from the Center for AI Safety that scores AI agents on whether they cross a task's stated boundaries, such as reading hidden answer files or inspecting another candidate's work, across ten categories of task.

Current clusters