Traditional evaluation of tool‑using language model agents focuses on whether the user’s task is completed, overlooking a common failure mode: the agent may finish the request while performing unauthorized side‑effects such as writing to the wrong object, expanding the request scope, or contacting an unintended recipient. We term this failure Collateral Damage (CD).\ \ To address this, we introduce SABench (Side‑effect‑Aware Benchmark), comprising 201 email, calendar, and ticket tasks. Each task combines one of four state structures (independent, dependent, externally reachable, cross‑tool) with one of four trap families (ambiguity, mis‑targeting, cascade harm, overreach) and provides a deterministic trace‑based oracle for both goal satisfaction and authorization compliance.\ \ Across 2,211 model‑task pairs from 11 frontier LLMs spanning major commercial and open‑source families, the overall CD rate is 15.2%. Ambiguity traps exhibit a 38.8% violation rate, while non‑ambiguous traps average 8.4%. Among runs that meet the goal, 12.8% still produce unauthorized writes. Cross‑tool and dependent states show roughly twice the CD rate of independent states. CD rates vary by 4.7× across models at comparable goal rates, indicating that task success alone is a poor proxy for safe tool use.\ \ The SABench task corpus is released at https://huggingface.co/datasets/Phaedrus/sabench; traces, oracle, code, and analysis pipeline are available at https://github.com/hjj194/sabench-artifact.\ \ Review