ATBench-Claw and ATBench-CodeX: Domain-Specific Agent Safety Benchmarks for OpenClaw and OpenAI Codex

arXiv·medium signal

ATBench extends its agent trajectory safety evaluation framework to OpenClaw and OpenAI Codex runtime environments. The benchmarks customize safety evaluation and diagnosis for each domain setting, enabling trajectory-level analysis of agent safety rather than single-turn evaluation. As agents move into increasingly diverse execution environments, trajectory-level safety benchmarks are essential for catching failure modes that single-turn tests miss.