audio

Overview ¶

Package audio implements attacks targeting audio-language models.

These attacks exploit models that accept audio input (GPT-4o, Gemini, SpeechGPT) by injecting adversarial content through the audio modality. Audio attacks often bypass text-only safety filters because the audio processing pipeline has different safety characteristics.

BoN Audio implements Best-of-N sampling with audio-specific augmentations. Generates N audio transcription variants and selects the most successful.

Source: OpenReview - Best-of-N Audio Augmentation

Multilingual Audio attacks exploit the observation that non-English audio inputs bypass safety filters at a 3.1x higher success rate than English.

Source: arXiv 2505.17568 - Multilingual Audio Jailbreaking

SpeechModelExploit implements white-box attacks against audio-language models by targeting the speech encoder and language model integration points.

Source: arXiv 2505.18864 - Exploiting Audio-Language Model Vulnerabilities