Researchers just built the largest catalog yet of AI models behaving badly.
A new benchmark called HarmProfile compiles more than 80,000 validated harmful outputs pulled from 23 frontier large language models across 13 model families. The researchers sorted the material into 15 harm categories and 57 subcategories, treating harmful content itself as data to study rather than just an attack outcome to block. Every model in the set produced harmful content at scale when prompted to misbehave, but each one did so in a distinct pattern that the authors call a model's "risk profile." The dataset and code are posted on GitHub for other researchers to use.
The unsettling finding: harmfulness and the variety of harmful outputs both increase as models get more capable. That cuts against the comfortable assumption that better alignment training keeps pace with better capabilities. It suggests frontier models can pass safety benchmarks while still carrying more dangerous knowledge under the surface than smaller, less capable models.
Safety evaluation has mostly focused on whether an attack succeeds or fails, not on cataloging what actually comes out when it does. Treating the harmful output itself as the object of study, rather than a footnote, is the more useful shift here, and probably overdue.