| --- |
| pretty_name: AudioAgentSecurity |
| tags: |
| - audio |
| - speech |
| - llm-agents |
| - agent-security |
| - prompt-injection |
| - red-teaming |
| - benchmark |
| - ai-safety |
| language: |
| - en |
| --- |
| |
| # AudioAgentSecurity |
|
|
| **The first comprehensive benchmark for audio instruction injection attacks against multimodal LLM agents.** |
|
|
| AudioAgentSecurity is the companion dataset of the paper |
| [*Piggybacking on Perception: Stealthy Concurrent Audio Prompt Injections against Multimodal LLM Agents*](https://arxiv.org/abs/2607.28165). |
| It supports the study of **concurrent audio prompt injection**: malicious audio instructions that |
| imperceptibly "piggyback" onto user speech in continuous audio interaction, hijacking multimodal |
| agents into executing malicious actions while the user keeps talking. |
|
|
| ## Dataset Summary |
|
|
| - **Task scenarios:** 8 real-world agent task scenarios |
| - **Attack patterns:** 10 distinct audio concealment / injection patterns |
| - **Evaluation (from the paper):** 11 state-of-the-art multimodal agents, including |
| **Gemini 3 Pro** and **GPT-4o-audio**; average Attack Success Rate (ASR) of **69.10%** against |
| Gemini 3 Pro |
| - **Defense:** the paper's **CADV** (Cascaded Audio Decoupling and Verification) mechanism, based on |
| acoustic source separation and cross-modal consistency analysis, achieves **>90% detection |
| success** across diverse attack vectors |
| - **Real-world validation:** human-volunteer experiments on the **Doubao AI Smartphone** in dynamic |
| real-world scenarios confirm high attack stealth and defense effectiveness; the vulnerability was |
| **responsibly disclosed to ByteDance** |
|
|
| ## Attack Technique |
|
|
| The benchmark instantiates the paper's two core techniques: |
|
|
| - **Instruction augmentation** — strengthening malicious audio instructions so they survive |
| concurrent mixing with user speech and remain executable by the agent; |
| - **Scenario concealment** — hiding malicious instructions inside perceptually plausible acoustic |
| carriers so they are imperceptible to users and hard to separate from benign speech. |
|
|
| ## Task Scenarios (8) |
|
|
| Benign user speech is drawn from eight real-world agent task scenarios |
| (`benchmark_dataset_mixed/benign/`, file-name prefix = scenario): |
|
|
| | Prefix | Scenario | |
| |---|---| |
| | `development_` | Software development assistance | |
| | `finance_` | Finance & banking services | |
| | `iot_` | IoT / smart-device control | |
| | `media_creative_` | Media & creative tasks | |
| | `navigation_` | Navigation & travel | |
| | `personal_assistant_` | Personal assistant tasks | |
| | `productivity_` | Productivity & office work | |
| | `system_control_` | System control operations | |
|
|
| ## Attack Patterns (10) |
|
|
| Concurrent attack audio is organized by concealment pattern |
| (`benchmark_dataset_mixed/mixed/<pattern>/`): |
|
|
| | Pattern | Description | |
| |---|---| |
| | `dialect/` | Attack speech disguised in regional dialects | |
| | `dolphin/` | Attack embedded in dolphin-like ultrasonic carriers | |
| | `foreign/` | Attack speech in foreign languages | |
| | `high_freq/` | High-frequency spectral concealment | |
| | `pulse/` | Pulsed / intermittent acoustic masking | |
| | `spectral_inversion/` | Spectrally inverted speech | |
| | `spectral_scramble/` | Spectrally scrambled speech | |
| | `speed/` | Time-stretched (speed-altered) speech | |
| | `texture/` | Attack blended into environmental sound textures | |
| | `whisper/` | Whispered attack speech | |
|
|
| ## Repository Structure |
|
|
| ``` |
| benchmark_dataset_mixed/ |
| ├── benign/ # clean user-speech audio (WAV), named <scenario>_<id>.wav |
| ├── mixed/ # concurrent audio (user speech + attack), by attack pattern |
| │ └── <pattern>/ # one of the 10 attack patterns above |
| ├── benign_text_dataset.json # benign user instruction texts |
| ├── malicious_text_dataset.json # malicious instruction texts used in attacks |
| └── metadata.json # record-level metadata: scenario / pattern / pairing / labels |
| ``` |
|
|
| Each *mixed* sample pairs a benign user utterance with a concealed malicious instruction, enabling |
| evaluation of agents under realistic concurrent-audio conditions; *benign* samples serve as the |
| no-attack control. |
|
|
| ## Intended Uses |
|
|
| - **Benchmarking multimodal LLM agents** against concurrent audio prompt injection. |
| - **Evaluating and training defenses** (e.g., source-separation and cross-modal consistency |
| based detection such as CADV). |
| - **Research on the audio attack surface** of voice-interactive agents, smartphones, and IoT |
| assistants. |
|
|
| This dataset is intended for **defensive security research and education**. |
|
|
| ## Ethical Considerations |
|
|
| - Benign speech is drawn from public speech corpora; malicious instruction texts are synthetic |
| research artifacts. The dataset contains no real user credentials or personal data. |
| - Attack audio is provided to study and defend against this threat class. |
| **Do not use it to attack real products or production voice assistants.** |
| - Vulnerabilities discovered in real products during the associated research were reported to the |
| affected vendor (ByteDance) following responsible disclosure practices. |
|
|
| ## Citation |
|
|
| If you use AudioAgentSecurity, please cite: |
|
|
| ```bibtex |
| @misc{liu2026piggybacking, |
| title={Piggybacking on Perception: Stealthy Concurrent Audio Prompt Injections against Multimodal LLM Agents}, |
| author={Mingxiao Liu and Yitong Li and Haoren Zhao and Yaoxiang Bian and Jianan Ma and Jian Zhang and Jialuo Chen and Xinhao Deng and Zhen Wang}, |
| year={2026}, |
| eprint={2607.28165}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.CR}, |
| url={https://arxiv.org/abs/2607.28165} |
| } |
| ``` |
|
|
| ## Contact |
|
|
| - Hugging Face: [Limax11](https://huggingface.co/Limax11) |
| - Issues and questions: please open a discussion on the dataset page. |
|
|
