AgentHazard: A Benchmark for Evaluating Harmful Behavior in Computer-Use Agents

TL;DR

AgentHazard is a comprehensive benchmark designed to evaluate the potential for harmful behaviors in computer-use agents, highlighting current vulnerabilities despite model alignment efforts.

Contribution

This paper introduces AgentHazard, a new benchmark with 2,653 instances to assess safety risks in autonomous agents across diverse attack strategies and operational scenarios.

Findings

Current systems show high vulnerability, with Qwen3-Coder powered Claude Code having a 73.63% attack success rate.

Agents often fail to recognize and interrupt harmful behaviors emerging from complex, multi-step sequences.

Model alignment alone does not ensure safety in autonomous agent systems.

Abstract

Computer-use agents extend language models from text generation to persistent action over tools, files, and execution environments. Unlike chat systems, they maintain state across interactions and translate intermediate outputs into concrete actions. This creates a distinct safety challenge in that harmful behavior may emerge through sequences of individually plausible steps, including intermediate actions that appear locally acceptable but collectively lead to unauthorized actions. We present \textbf{AgentHazard}, a benchmark for evaluating harmful behavior in computer-use agents. AgentHazard contains \textbf{2,653} instances spanning diverse risk categories and attack strategies. Each instance pairs a harmful objective with a sequence of operational steps that are locally legitimate but jointly induce unsafe behavior. The benchmark evaluates whether agents can recognize and interrupt harm arising from accumulated context, repeated tool use, intermediate actions, and dependencies across steps. We evaluate AgentHazard on Claude Code, OpenClaw, and IFlow using mostly open or openly deployable models from the Qwen3, Kimi, GLM, and DeepSeek families. Our experimental results indicate that current systems remain highly vulnerable. In particular, when powered by Qwen3-Coder, Claude Code exhibits an attack success rate of \textbf{73.63\%}, suggesting that model alignment alone does not reliably guarantee the safety of autonomous agents.