DecodingTrust Agent Platform
A Real-World Simulation Platform for
Advanced Red-Teaming of AI Agents
Powered by DT-Red, our autonomous red-teaming agent, and DT-Bench, a comprehensive benchmark with 30+ sandbox environments, 15+ domains, and 500+ tasks per domain.
A research collaboration • Paper available on arXiv
Comprehensive Security Evaluation
Built for researchers and practitioners to rigorously test AI agent security across real-world regulatory scenarios.
High-Fidelity Sandboxes
30+ realistic environments including Gmail, PayPal, Databricks across finance, healthcare, and e-commerce.
Policy-Aligned Evaluation
Risks derived from domain-specific policies like FINRA in Finance and Salesforce AI Use Policy for regulatory compliance.
DT-Red: Autonomous Red-Team Agent
First autonomous agent that iteratively optimizes attack vectors and injection locations to uncover vulnerabilities.
Black-Box Evaluation
Unified protocol supporting evaluation of any agentic system including Claude Code, Cursor, and custom agents.
Comprehensive Task Coverage
Over 500 benign and malicious tasks per domain ensuring thorough security evaluation across attack surfaces.
Scalable Discovery
Efficiently discover diverse, policy-aligned attack vectors with high success rates through automated optimization.
Spanning 15+ Real-World Domains
Each domain includes policy-aligned evaluation scenarios based on actual regulatory and compliance requirements.
Featured Sandbox Environments
Security Robustness Leaderboard
Defense rates for top AI agents on DT-Bench v1.0