ATBench: Agent Trajectory Safety Benchmark
What it is
ATBench is a family of trajectory-level safety benchmarks designed to evaluate long-horizon, tool-using AI agents. It provides a dataset of 1,000 realistic agent trajectories (with a legacy 500-case release) that are labeled as safe or unsafe to test the ability of models and guardrails to detect safety violations in complex, multi-step agent interactions.
The problem it solves
Existing safety benchmarks often lack the complexity, diversity, and realistic tool-usage patterns found in real-world long-horizon agent trajectories, making it difficult to accurately assess agent safety.
What you could use it for
- Evaluating the safety performance of large language models on agent tasks
- Benchmarking specialized safety guard models (e.g., AgentDoG, LlamaGuard)
- Diagnosing specific types of safety failures in tool-using agents
- Training fine-tuned models for agent trajectory safety classification
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.