OmniTrust is a comprehensive security platform designed to evaluate the trustworthiness of large language models (LLMs) and multimodal models. Supporting both white-box and black-box LLMs, OmniTrust offers an out-of-the-box toolkit that is intuitive and easy to use—allowing developers and researchers to run complex security evaluations with a single line of code.

At its core, OmniTrust comprises six powerful modules that cover key areas of model integrity and safety:

1. Safety Module: Protects against a wide array of attack techniques, including AutoDAN, CipherChat, and more, enabling both attack and defense evaluation on LLMs and multimodal models.
2. Privacy Module: Focuses on fine-tuned models to detect data extraction, membership inference, and prompt word stealing attacks, ensuring compliance with privacy standards.
3. Detectability Module: Provides watermarking addition, detection, and evaluation, safeguarding against tampering and ensuring content integrity.
4. Truthfulness Module: Evaluates knowledge consistency and reasoning accuracy across tasks such as Q&A, code generation, and cross-modal content generation, mitigating the risk of misinformation.
5. Fidelity Module: Protects LLMs from prompt injection attacks, ensuring consistent and fair scoring mechanisms for judge-based tasks.
6. Fairness Module: Analyzes and mitigates bias in LLMs, providing insights into potential issues such as egocentric or attentional bias, and offering solutions for effective calibration.

OmniTrust comes with a wide range of datasets tailored for each module, along with complete metrics to evaluate the overall trustworthiness of a model. The platform’s simplicity is key: a single line of code allows users to run any module, choose models, load datasets, and generate detailed, actionable reports—making it a powerful and user-friendly solution for securing and evaluating generative foundation models.