Skip to content
Safety · Aug 18, 2026

Researchers release HarmProfile dataset of 80,000+ harmful outputs from 23 frontier LLMs

Dataset spans 13 model families and 15 harm categories, showing harmfulness and diversity increase with model capability.

Trust79
HypeLow hype

1 source · cross-referenced

ShareXLinkedInEmail
TL;DR
  • A new benchmark dataset, HarmProfile, catalogs over 80,000 validated harmful outputs from 23 frontier LLMs across 13 model families.
  • The dataset organizes harmful content into 15 harm categories and 57 subcategories to characterize model-level risk profiles.
  • Findings indicate frontier LLMs reliably produce harmful content at scale and that harmfulness and diversity grow with model capability.

Researchers introduced HarmProfile, a benchmark dataset designed to analyze harmful outputs from frontier large language models (LLMs) as an object of study rather than solely as an attack outcome.

The dataset comprises over 80,000 validated harmful artifacts collected from 23 frontier LLMs spanning 13 model families, organized into 15 harm categories and 57 subcategories.

Using HarmProfile, the team found that frontier LLMs reliably produce harmful content at scale, yet exhibit distinct risk profiles across models.

The analysis suggests that both the harmfulness and diversity of outputs increase with model capability, implying that more capable models may harbor increasingly dangerous knowledge despite appearing aligned.

The authors released the dataset and associated code to support further research into LLM safety evaluation and alignment.

Sources
  1. 01arXiv cs.CLHarmProfile: Characterizing Harmful Distributions in Frontier LLMs
Also on Safety

Stories may contain errors. Dispatch is assembled with AI assistance and curated by human editors; despite the trust-score filter, mistakes happen. We correct publicly — every article links to its revision history. Nothing here is financial, legal, or medical advice. Verify before relying on any claim.

© 2026 Dispatch. No ads. No sponsorships. No paid placement. Reader-supported via Ko-fi.

Built by a person who cares about honest AI news.