HarmProfile: Characterizing Harmful Distributions in Frontier LLMs
A new benchmark exposes the hidden diversity and severity of safety failures across frontier LLMs.
Frontier large language models (LLMs) safety evaluation has largely treated harmful generation as an attack outcome rather than as an object of analysis. Consequently, little is known about the harmful outputs produced during model misbehavior, partly because large-scale, high-qu...
ArXiv CS.CL
Source
Generated insight by OnAirToday AI
HarmProfile shifts the focus from simple adversarial testing to a comprehensive analysis of model-level risk profiles, revealing that increasing model capability often hides deeper, more dangerous vulnerabilities. By quantifying the nuances of safety failures across 80,000 artifacts, this research provides a standardized framework to understand how frontier LLMs actually harbor harmful knowledge.
Comments
Keep comments relevant and respectful. No links, advertising, self-promotion, or spam.