openrt 0.0.1

An Open-Source Red Teaming Framework for Large Language Models (LLMs) and Vision-Language Models (VLMs)

OpenRT is a comprehensive framework for red teaming LLMs and VLMs with 30+ attack methods.

Features:

• 30+ attack methods (black-box and white-box)
• Multi-modal support (text and image attacks)
• Modular plugin architecture
• Configuration-driven experiments
• Multiple evaluation strategies
• Registry-based component loading

The framework follows a modular design with:

• Attack implementations (AutoDAN, GeneticAttack, DeepInception, etc.)
• Model integrations (OpenAI API, custom base_url support)
• Dataset handlers (StaticDataset, JSONLDataset)
• Evaluators (KeywordEvaluator, LLMEvaluator)
• Strategy patterns (Advancers, Propagators, Judges)