When Robots Mishear Us: Mapping the Safety Risks of Voice-Controlled Embodied AI
- ID:
a534e3a7 - 原文链接: https://arxiv.org/abs/2608.28518
- PDF: https://arxiv.org/pdf/2608.28518v1
- 作者: Sihan Jia, Oliver Lemon
- 发布日期: 2026-08-28
- 更新日期: 2026-08-28
- 主分类: cs.AI
- 分类: cs.AI, cs.CL, cs.RO
- 备注: N/A
- 归档分类: agents
- 标签: embodied-ai、asr、safety、voice-interface、benchmark
- 质量评分: 4/5
- 抓取时间: 2026-09-01T04:24:13Z
中文导读
研究语音控制的具身智能（EAI）中 ASR 识别错误是否会导向不安全输出。作者模拟 ASR 错误并与现有安全基准 SafeAgentBench、POEX 结合评估：部分错误类型保留语义结构但放大有害歧义，另一些则削弱模型拒绝行为，使不安全计划得以生成并执行；自动纠错在某些场景能降低风险，但并不总是有效。结论是 ASR 错误对语音控制具身 AI 构成显著安全风险，为语音入口 agent 执行链的安全评测提供了新的攻击面视角。
为什么值得关注
语音入口不是可信通道：ASR 错误在 SafeAgentBench/POEX 上既放大了有害歧义，又削弱拒绝行为，语音入口的具身 agent 应当被纳入对抗性安全评测。
关键信息
- 论文标题：When Robots Mishear Us: Mapping the Safety Risks of Voice-Controlled Embodied AI
- 作者：Sihan Jia, Oliver Lemon
- arXiv：https://arxiv.org/abs/2608.28518
- PDF：https://arxiv.org/pdf/2608.28518v1
- 发布时间：2026-08-28
- 更新日期：2026-08-28
- arXiv 分类：cs.AI, cs.CL, cs.RO
- 备注：N/A
- 关联标签：embodied-ai、asr、safety、voice-interface、benchmark
English Abstract
We investigate whether automatic speech recognition (ASR) errors in user input can lead to unsafe outputs from Embodied AI (EAI) models. We find that ASR errors can lead to harmful instructions being accepted and executed by EAI models, thereby reducing safety. We simulate ASR errors and combine them with existing safety benchmarks (SafeAgentBench and POEX) to evaluate how different errors affect embodied AI safety. We find that some of them preserve semantic structure but increase harmful ambiguity, while others weaken the model refusal behaviour and allow unsafe plans to be generated and executed. We show that in some cases automatic correction of ASR errors can reduce the risk, but this is not always effective. Overall, we show that ASR errors lead to significant safety risks for embodied AI.
English Summary
The authors investigate whether automatic speech recognition (ASR) errors in user input lead to unsafe outputs from embodied AI (EAI) models. Simulated ASR errors are combined with existing safety benchmarks (SafeAgentBench and POEX) to evaluate how error types affect embodied AI safety. Some error classes preserve semantic structure while increasing harmful ambiguity; others weaken refusal behaviour and allow unsafe plans to be generated and executed. Automatic correction of ASR errors reduces risk in some cases but is not always effective. Overall, ASR errors introduce significant safety risks for voice-controlled embodied AI.
Obsidian Notes
- 内容由
opencli arxiv paper拉取 arXiv 元数据与摘要生成。 - 中文导读与价值判断均锚定在条目已有摘要、论文摘要、作者、日期与分类信息上；未补充论文摘要之外的实验细节。
- 抓取时间戳：2026-09-01T04:24:13Z（UTC）。