About DT-Agent
DecodingTrust Agent — a real-world simulation platform for advanced red-teaming of AI agents.
Our Mission
DT-Agent is a real-world simulation platform for advanced red-teaming of AI agents. Our mission is to provide researchers and practitioners with the tools needed to rigorously evaluate AI agent security across diverse, real-world domains.
The platform encompasses 30+ high-fidelity sandbox environments (e.g., Gmail, PayPal, Databricks) spanning over 15 domains, with over 500 benign and malicious tasks per domain. All evaluated risks are derived directly from domain-specific safety and security policies such as FINRA in Finance and the Salesforce AI Use Policy.
Research Team
Researcher 1
Lead Researcher
University A
Researcher 2
Security Expert
University B
Researcher 3
ML Engineer
Company C
Researcher 4
Benchmark Design
University A
Affiliations
Contact
For research collaborations, questions, or media inquiries, please reach out:
Citation
@article{dt-agent2025,
title={DT-Agent: A Real-World Simulation Platform for
Advanced Red-Teaming of AI Agents},
author={Author1 and Author2 and Author3 and Author4},
journal={arXiv preprint arXiv:2501.xxxxx},
year={2025}
}