Our take
- Activity
- 30
- Community
- 66
- Issues
- 46
- Pull requests
- 46
+22 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.
Stars
946Forks
95Change since Nov 14, 2026: 0 stars·0 forks
Overview
EasyJailbreak is a Python framework for constructing adversarial prompts and evaluating jailbreak attempts against large language models. It organizes attacks as a loop of prompt selection, mutation, constraint checks, model inference, and evaluation, rather than tying experiments to a single attack recipe.
It is intended for LLM security research and development. Use it to run included attack recipes or combine reusable components when comparing attack approaches and model behavior. The project links to a paper, documentation, and a website with examples and results.
Key Features
- Implements multiple attack recipes, including PAIR, GCG, GPTFuzz, AutoDAN, and TAP.
- Provides reusable selectors, prompt mutators, constraints, and evaluators.
- Supports Hugging Face models and several cloud-hosted model integrations described in the README.
- Loads jailbreak datasets from the project’s online dataset collection or local files.
- Produces reports containing prompts, target-model responses, and evaluation scores.
- Offers a component-based path for researchers to assemble custom attacks.
Use Cases
- LLM security researchers can compare implemented attack recipes against selected target models.
- Model developers can assess how a model responds to adversarial prompts during security testing.
- Framework contributors can implement new selectors, mutations, constraints, or evaluators.
- Students and practitioners can explore published jailbreak methods using a common workflow and dataset interface.
Getting Started
Python 3.9 or newer is required. Install the package with:
pip install easyjailbreak
For development or custom components, see the README and documentation.
Considerations
- This is a security research and evaluation framework, not a general-purpose chat or model-serving tool.
- Running attacks requires access to target models and, depending on the recipe, may also require attack or evaluation models and API credentials.
- Local model use can require substantial compute resources; the README does not specify hardware requirements.
- The GPL-3.0 license is a consideration for projects that intend to incorporate or redistribute the code.
Found this useful?
Share it with someone who would like EasyJailbreak.