FORTRESS: Results as Published

FORTRESS tests whether a model stays safe on national-security and public-safety prompts and whether it still answers matched harmless prompts. The public set has 500 expert-written adversarial prompts across chemical, biological, radiological, nuclear and explosive hazards, political violence and terrorism, and criminal and financial illicit activity, with 10 subcategories in all. Each prompt has a rubric of 4 to 7 yes-or-no questions, plus a benign twin used to check over-refusal. The paper scores an average risk score from the adversarial rubrics and an over-refusal score from the benign prompts.

Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.