Yunhao Feng, Yifan Ding, Yingshui Tan, Yige Li, Yige Li, Yutao Wu, Yifeng Gao, Kun Zhai, Xingjun Ma · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3838633
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
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: harmful behavior may emerge through sequences of individually plausible steps that collectively lead to unauthorized operations, sensitive data access, or destructive execution. We present AgentHazard , a benchmark for evaluating harmful behavior in computer-use agents. AgentHazard contains 2,653 instances spanning 10 risk categories and 10 attack strategies. Each instance pairs a harmful objective with a sequence of operational steps that appear locally legitimate but jointly induce unsafe behavior. The benchmark evaluates whether agents can recognize and interrupt harm arising from accumulated context and repeated tool use across multi-step trajectories. We evaluate AgentHazard on Claude Code, OpenClaw, and IFlow using diverse open and openly deployable models from the Qwen, Kimi, and GLM families. Our experimental results indicate that current systems remain highly vulnerable: when powered by Qwen3-Coder, Claude Code exhibits an attack success rate of 73.63% , suggesting that model-level alignment does not reliably guarantee agent-level safety. To facilitate future research, we open-source the dataset, supplementary materials, and benchmark construction pipeline at https://yunhao-feng.github.io/AgentHazard/.
Last synced
This work has 20 recorded citations, but citing papers have not been linked locally yet.
No comments yet — start the discussion below.