Voice Jailbreak Attacks Against GPT-4o
Paper • 2405.19103 • Published
Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains synthesised speech of jailbreak attack prompts (fraud, hate speech, illegal activity, physical harm, pornography, and privacy violation instructions). Request access only if you have a legitimate safety research or model evaluation use for it.
Log in or Sign Up to review the conditions and access this dataset content.
Voice Jailbreak Attacks — synthesised audio
Pre-synthesised speech for Voice Jailbreak Attacks Against GPT-4o
(Shen et al., 2024), spoken with OpenAI
text-to-speech (tts-1, voice fable) from the forbidden questions and text
jailbreak templates published in the paper's
repository.
Content warning: this dataset contains spoken instructions for fraud, hate speech, illegal activity, physical harm, pornography, and privacy violations, and jailbreak templates written to elicit exactly that from an AI model. Access is gated; request access only if you have a legitimate research or evaluation use for it.
What's here
Two subsets, each an audiofolder-compatible
directory (wav files plus a metadata.csv with a file_name column):
| subset | clips | source text | typical clip length |
|---|---|---|---|
baseline |
30 | the bare forbidden question | ~2-6 seconds |
textjailbreak |
150 | each question wrapped in one of 5 jailbreak templates | ~30 seconds - 6 minutes |
metadata.csv columns: file_name, sample_id, subset, question_id,
q_id, content_policy_id, content_policy_name, question,
jailbreak_id, jailbreak_community, prompt (the exact text spoken).
Used by the voice_jailbreak_attacks
Inspect AI eval, which downloads this dataset
by default instead of re-synthesising speech on every machine. The eval can
still synthesise from source on demand (-T source=synthesize); this repo
exists so that most users don't have to.
Why the audio wasn't published before
Neither the paper nor HELM's VoiceJailbreakAttacksScenario,
which the eval ports, distributes audio; both leave synthesis to a manual run
of the paper's tts/prompt2audio.py. This repo runs that step once, so it
doesn't have to be repeated (in cost or wall-clock) by everyone who wants to
run the eval.
Provenance and licensing
- Source text (questions and jailbreak templates): MIT licensed by TrustAIRLab, from the paper's repository.
- Audio: synthesised with the OpenAI API under an account that owns the output (OpenAI's Terms of Use assign the caller "all right, title, and interest in and to Output"), so republishing it here does not require OpenAI's separate permission. It is offered under the same MIT terms as the source text.
This dataset is gated, not because of a licensing restriction, but because the content itself is a set of ready-to-use jailbreak attacks — gating adds friction against casual/automated misuse without blocking legitimate safety research and evaluation, which is what it's for.
Citation
@article{SWBZ24,
author = {Xinyue Shen and Yixin Wu and Michael Backes and Yang Zhang},
title = {{Voice Jailbreak Attacks Against GPT-4o}},
journal = {{CoRR abs/2405.19103}},
year = {2024}
}
- Downloads last month
- 32