DecodingTrust-Agent Platform documentation
DTap is a full-stack simulation world for AI agent red-teaming, simulating 50+ realistic environments across 14 high-stakes domains with diverse attack injection interfaces to enable scalable, reproducible, and transferable evaluation of agent security under realistic multi-surface attacks.
Quick Start
Run your first benchmark in 5 minutes.
Install from Source
Clone, set up the venv, install editable.
Eval with decodingtrust-agent
Run a JSONL of tasks across 14 domains.
Off-the-Shelf Agents
OpenAI, Claude, Google, LangChain, OpenClaw.
Browse by domain
Browse environments
Each environment ships intro prose, GUI screenshots, MCP tool inventory, and policy-aligned attack categories. Open the Environment section in the sidebar.