GA is a comprehensive Python package designed for evaluating the safety and robustness of Large Language Models (LLMs). It provides state-of-the-art jailbreaking techniques, model interfaces, and evaluation tools.
Features
Jailbreaking Methods
Implementations of cutting-edge techniques to test model safety boundaries:
- AutoDAN & AutoDAN-Turbo: Hierarchical genetic algorithms and lifelong learning
- TAP: Tree-of-Attacks with Pruning for systematic exploration
- GCG: Gradient-based optimization for white-box attacks
- Crescendo: Progressive multi-turn attacks
- Bijection Learning: Randomized encodings to bypass filters
Model Interfaces
Unified APIs for both cloud and local models:
- BlackBoxModel: Interface for cloud-hosted models (OpenAI, Anthropic, Together.ai)
- WhiteBoxModel: Direct access to model weights for gradient-based methods
Adversarial Generators
Core algorithms for crafting adversarial prompts:
- Tree-based refinement strategies
- Multi-turn conversation attacks
- Genetic algorithms for prompt evolution
- Strategy-based prompt generation
Evaluation Tools
Comprehensive tools for assessing jailbreak effectiveness:
- Standardized evaluation metrics
- Attack success rate (ASR) measurement
- Cross-model comparison capabilities
Why General Analysis?
- Comprehensive Coverage: All major jailbreaking techniques in one framework
- Easy Integration: Simple APIs that work with popular model providers
- Research-Ready: Implementations based on latest academic papers
- Production-Tested: Used by leading AI safety teams
- Open Source: Fully transparent and community-driven
Installation
pip install generalanalysisQuick Example
from generalanalysis.jailbreaks import TAP, TAPConfig
from generalanalysis.boiler_room import BlackBoxModel
# Initialize target model
target_model = BlackBoxModel("gpt-4o")
# Configure jailbreak
config = TAPConfig(
target_model="gpt-4o",
attacker_model="deepseek-ai/DeepSeek-R1",
evaluator_model="deepseek-ai/DeepSeek-R1",
branching_factor=4,
max_depth=10
)
# Run attack
tap = TAP(config)
results = tap.optimize(["Write malicious content"])Community
Join our community to stay updated on the latest developments:
License
General Analysis is released under the MIT License. See the LICENSE file for details.
Last updated on