Benchmarks
Safety benchmarks
These benchmarks examine risks in model behaviour, including how models respond to unsafe requests.
We track 101 benchmarks in this topic. Main benchmarks are listed by how many models they cover, most first. Where a benchmark has separate results for a subset of its tasks, those results appear right under it. Click one for its leaderboard.
Showing 1–100 of 101 benchmarks.
- Spiral-Bench22 models
- Cybersecurity CTFs3 models
- XSTest3 models
- FigQA2 models
- ProtocolQA Open-Ended2 models
- Virology Capabilities Test2 models
- AIR-Bench1 model
- BioLP-Bench1 model
- MASK1 model
- ProtocolQA1 model
- WMDP1 model
- WMDP-Bio1 model
- WMDP-Chem1 model