AI Agent Safety Benchmarks Reveal Deep Inconsistencies: New Study

Researchers led by Miles Q. Li have published the first systematic analysis dedicated to safety benchmarks for LLM-based autonomous agents, cataloging 40 behavioral benchmarks from 2023 to 2026 plus 5 adjacent artifacts. The team proposes a six-axis taxonomy of benchmark evaluation methodology, revealing a landscape where the choice of benchmark can produce contradictory safety conclusions. Their coverage matrix shows broad risk coverage but limited methodological convergence, with the core of behavioral benchmarks concentrated in sandboxed, constrained, and often safety-only evaluations. Environment fidelity systematically shapes reported safety, and the field disproportionately tests externally imposed risks over agent-internal ones.

The study's cross-benchmark consistency check using Kendall's W concordance analysis found no evidence of ranking concordance across evaluation dimensions (W=0.10, p=0.94), indicating that different benchmarks rank agents' safety inconsistently. Metric fragmentation makes comparison nearly impossible, and robustness remains effectively unbenchmarked. The authors release structured metadata, full taxonomy codings, risk annotations, and all experimental artifacts to enable replication. They also propose minimum reporting standards for future benchmarks, addressing the urgent need for standardized evaluation as AI agents are deployed in high-stakes environments.