Cybersecurity
SecProbe
Adaptive evaluation of coding agents on repository-scale vulnerability repair, selecting or generating tasks to reveal capability gaps.
Evaluations
Make progress measurable.
Open benchmarks that reveal where models fall short, and where improvement begins.
Professional work
JobBench
Evaluating agents on real-world professional tasks that people want to delegate to AI.
Interactive narrative
NARRA-Gym
An executable evaluation environment for agents that sustain interactive stories, adapt to users, and manage narrative state across multiple turns.