HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal
by Mantas Mazeika, Long Phan and colleagues
Study a standardized approach to red-team evaluation and robust refusal without copying attack recipes.
HarmBench defines an evaluation framework for automated red teaming and compares attack methods, models and defenses. The paper also studies adversarial training for robust refusal; the index retains only descriptive research context.
- Level
- Advanced
- Models
- model-agnostic
- Category
- Jailbreaking and red teaming
- Tags
- primary-source, jailbreak-research
- Source date
- 2024-02-06
- Added
- 2026-10-18
- Last verified
- 2026-10-18
- Status
- ok
- Prerequisites
- Requires familiarity with LLM evaluation, adversarial robustness and experimental design.
- Subcategory
- jailbreak research
- Region
- West
- Original language
- en
- Access
- free
- Resource reuse
- open: The inspected paper HTML explicitly states License: CC BY 4.0; the reuse claim is scoped to this paper. Facet definitions
Summary in our own words. Rights remain with the original creator. Credit and correction policy.