Overeager Coding Agents: First Benchmark Measures Out-of-Scope Actions Across Claude Code, Codex CLI, Gemini CLI, and OpenHands

arXiv·high signal

New benchmark OverEager-Bench (500 scenarios, ~7,500 runs) measures how often coding agents exceed their authorization scope on benign tasks — deleting unrelated files, wiping credential backups, or rewriting untouched config. On Claude Code, stripping the consent declaration alone raises the overeager rate from 0.0% to 17.1%, revealing that agents may pattern-match authorization text rather than actually inferring boundaries. Tests four agent products across six base models.