Open-weight models you can run on your own servers, tested the same way each time. How durable are the guardrails, what does the model know that could cause harm, and what should a business do about it?
Refuses harmHarmBench: share of 200 harmful requests the original model refuses. Higher is safer.
After tamperingThe same test after we remove the refusal behavior from the weights with a published method. Low means anyone with a copy can switch the guardrails off.
Hazard knowledgeWMDP accuracy across biology, chemistry and cybersecurity. Random guessing scores 25%.
Refuses safeOR-Bench hard: safe questions that only sound risky. High means staff will hit refusals on legitimate work.
Every card uses the same tests and the same rating thresholds, so models can be compared directly. When the method changes, the version number on every card changes with it. We publish results, never stripped weights or removal recipes.