ThuCCSLab

JailbreakEval: Compare LLM Jailbreak Evaluators

JailbreakEval brings together automated methods for assessing whether language-model responses comply with jailbreak attempts. Researchers can compare evaluators across datasets, while developers can build and benchmark new evaluation methods.

View on GitHub Compare
Stars
199
Forks
13
Stars, last 30 days
+0

Our take

Health score

24/100

How it is scored
Activity
19
Community
34
Issues
21
Pull requests
20

+0 stars in the last 30 days

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

200

Forks

14

Change since Nov 14, 2026: 0 stars·0 forks

Overview

JailbreakEval is a Python toolkit for evaluating whether responses to harmful prompts indicate a successful jailbreak. It addresses the difficulty of manually reviewing large collections of attempts and the fact that different automated evaluators can produce different judgments.

Rather than imposing one universal scoring method, the toolkit brings multiple evaluator types into a common framework. It is suited to jailbreak researchers comparing methods and to developers implementing or benchmarking evaluators. See the project README and paper for details.

Key Features

  • Offers string-matching, chat-based, text-classification, and voting evaluators.
  • Includes presets based on evaluators used in published jailbreak and safety research.
  • Provides a Python API for assessing individual attempts.
  • Provides a CLI for evaluating CSV datasets with question and answer columns.
  • Reports evaluation coverage, success rates, timing, and token counts where applicable.
  • Compares automated results with supplied labels using agreement metrics.
  • Supports evaluator configuration and extension for custom evaluation needs.

Use Cases

  • Jailbreak researchers can compare how established evaluators score the same attack dataset.
  • Safety teams can run automated assessments over labeled collections instead of manually reviewing every response.
  • Evaluator developers can implement a new method within the framework and benchmark it against existing presets.
  • Researchers using different evaluation paradigms can examine how string matching, model-based judgments, or voting affect results.

Getting Started

Install the package from PyPI:

pip install JailbreakEval

The README covers Python usage, CLI evaluation, configuration, and available evaluator presets.

Considerations

  • Evaluators are not interchangeable or universally accurate. The project’s reported results show that performance varies by evaluator and dataset, so choose methods appropriate to the study and validate them against human labels when possible.
  • Some presets depend on external services such as OpenAI or Perspective API, or on Hugging Face models. These may require credentials, network access, and additional model resources.
  • Dataset-based CLI evaluation expects a UTF-8 CSV with at least question and answer columns. Optional labels enable comparison with annotations.
  • Automated scores are aids for analysis, not substitutes for carefully defined evaluation criteria or human review in high-stakes settings.

Found this useful?

Share it with someone who would like JailbreakEval.

Comparisons