AI red teaming & LLM evaluation
HarmBench is an open-source red teaming benchmark from Center for AI Safety. It is released under the MIT license and is offered as: CLI, Self-hosted.
Best for
Researchers who need a reproducible way to compare red teaming attacks and refusal defenses across many models
Consider instead if
HarmBench is an evaluation framework for automated red teaming of language models. The README describes it as a framework for evaluating red teaming methods and LLM attacks and defenses, and it cites 18 red teaming methods and 33 target LLMs and defenses. You run a pipeline of scripts that generates test cases, generates completions from the target model, and has a classifier label each completion; the classifiers are published on Hugging Face. It supports models loaded through transformers, several closed-source APIs and some multimodal models. The code is MIT-licensed on GitHub, and the last push to the repository was on 2024-08-16. It is a research benchmark you run yourself, not a scanner for production applications.
This summary is our own wording of the sourced facts on this page.
Each value links to the page that states it, with the date we checked it.
As documented by the vendor or project. The tag is our label from a controlled vocabulary.
Our editorial model, not a standard. The primary layer follows the category, except for tools whose main documented job is detection and response; other layers rest on a capability the sources document. How the model works.
| Layer | Role | How it applies | Based on |
|---|---|---|---|
| L4 Testing Testing and red teaming | Primary | Attack and evaluate models, apps and agents on purpose, before release and after every change, to find what the other layers miss. | Follows its category, AI red teaming & LLM evaluation |
The risk list is OWASP's (Top 10 for LLM Applications, 2025). The mapping is our analysis, and each row rests on a capability the sources document.
No OWASP LLM risk is mapped to this tool in our analysis.
More in AI red teaming & LLM evaluation, ranked within the category by Coverage Score.
NVIDIA
Probes language models and LLM apps for prompt injection, jailbreaks, data leakage and more.
Best for: Teams that want a free command-line scanner they can point at any model or LLM endpoint before release
Reviewed
Giskard AI
Generates adversarial test suites for LLM agents and RAG apps, runs them and reports failures by threat type.
Best for: Python teams that want a free library to scan an LLM agent or RAG app for prompt injection and harmful output, and to write pass or fail checks for it
Reviewed
OpenAI (Promptfoo)
Runs evals and red team scans against LLM apps from a CLI or CI pipeline and writes a vulnerability report.
Best for: Developer teams that want one open-source CLI for both prompt evals and red teaming, run locally and in CI
Reviewed
Yes. HarmBench is released under the MIT license, and the code is on GitHub at centerforaisafety/HarmBench.
It runs red teaming methods against target models and uses a classifier to label whether each completion shows the targeted behavior, so attack success rate can be computed.
The GitHub API shows the repository is not archived and was last pushed to on 2024-08-16.
HarmBench is released under the MIT license.
Source Source, checked October 21, 2026: github.com (opens in a new tab) Checked
HarmBench is offered as: CLI, Self-hosted.
Source Source, checked October 21, 2026: github.com (opens in a new tab) Checked
Pricing model for HarmBench: Free.
Source Source, checked October 21, 2026: github.com (opens in a new tab) Checked
HarmBench had 1,076 GitHub stars on October 21, 2026.
Source Source, checked October 21, 2026: api.github.com (opens in a new tab) Checked
We have no commercial relationship with Center for AI Safety.