Research
What We Learn by Breaking AI
We study how language models fail in practice and publish what we find. Our research shapes the tests we run for our clients.
-
Safety That Doesn't Translate: Harmful Prompts in All 24 EU Languages
We sent 1,053 harmful requests in every official EU language to 34 open-weight LLMs. Outside English they comply more than twice as often, and a model's English results say little about its safety in Latvian, Maltese or Irish. With an interactive explorer of the results by language, model and harm type.