DynaGuard

Dynamic guardian models (1.7B/4B/8B) that enforce user-defined natural-language policies instead of a fixed harm taxonomy, with optional chain-of-thought rationales and a fast no-rationale mode; beats GPT-4o-mini on policy-compliance benchmarks and ships with the DynaBench training/eval set. Tom Goldstein's group jointly with Capital One — note the industry co-attribution.