ATBench is a benchmark and diagnostic protocol for evaluating the safety of long-horizon, tool-using autonomous agents by analyzing complete execution traces, including user requests, agent actions, tool calls, and environment feedback. It uses a three-dimensional safety taxonomy—Risk Source, Failure Mode, Real-world Harm—to enable trajectory-level evaluation and fine-grained diagnosis of where risk enters, how it manifests, and what harm could follow. The release includes 1,000 audited trajectories (ATBench) with a legacy ATBench500 lineage, a diverse set of tools, and benchmark results across open- and closed-source models to illustrate gaps between general safety and trajectory-level safety.
AI named ATBench in September 2026.
Brand page for yuli-cs-net-2