OpenRT: An Open-Source Red Teaming Framework for Multimodal LLMs

🧩 Framework Innovation

A Modular, Extensible, and Composable Framework for Red Teaming. OpenRT decoupled red-teaming into five orthogonal dimensions: models, datasets, attacks, judges, and metrics. Powered by a unified Component Registry, it enables seamless plug-and-play composition, spanning from simple prompts to complex multi-turn and multi-agent strategies. This standardized design decouples complex search and optimization strategies into infrastructure, enabling developers to reuse underlying search capabilities without redundant implementation.

🚀 Extreme Efficiency

30x Acceleration with One-Line Command. Designed for large-scale evaluation, OpenRT leverages a high-concurrency architecture based on AsyncIO and ThreadPool. By achieving dual parallelism in inference and scheduling, it boosts throughput by 30x compared to serial baselines. With just one line of command, users can execute a high-throughput scan covering the entire pipeline—from attack generation and auto-judgment to safety reporting.

🛡️ Most complete and Up-to-date

37+ SOTA Attack Paradigms (Continuously Updated). OpenRT features the industry's most extensive algorithm library, integrating 37+ state-of-the-art attack methods and continuously tracking the latest research advancements. Our arsenal covers the full threat spectrum, including multimodal attacks, multi-agent coordination, logic obfuscation, and iterative optimization. More than just a toolbox, OpenRT serves as a standardized infrastructure for the safety acceptance of next-generation frontier models.

@article{OpenRT2026,
  title={OpenRT: An Open-Source Red Teaming Framework for Multimodal LLMs},
  author={Shanghai AI Lab},
  journal={arXiv preprint arXiv:2601.01592},
  year={2026}
}

Contact

Corresponding email: tengyan@pjlab.org.cn

Project homepage: https://github.com/AI45Lab/OpenRT