The BaseResearcharXiv
AUTEUR: Acoustic Plan Learning for Red-Teaming Text-to-Audio Models
A preprint presents AUTEUR, an automated red-teaming method for text-to-audio models that learns structured acoustic plans to produce test prompts passing text screening but yielding harmful audio. The authors report attack success of 79.4% on TangoFlux and 95.6% on Kling.
Why it matters
The same thing, for different readers.
What it means for you
Researchers report that text screening alone does not stop audio-generation models from producing harmful sounds.
For researchers
Prompts transferred across open- and closed-source models, which the authors read as shared acoustic vulnerabilities.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- A plan generator trained with supervised fine-tuning then reinforcement learning on audio feedback, used to write and revise test prompts against four target text-to-audio systems.
- What it found
- The authors report attack success rates of 79.4% on TangoFlux and 95.6% on Kling, and say the prompts also transfer across open-source and closed-source models.
- What it does not show
- The abstract does not define how success was judged; the other two target systems are not named in the excerpt.
- Status
- A preprint: not yet peer reviewed
The record
- Questions
- Are AI companies putting stronger safeguards into practice?
- People
- Jiashu Zhang (Wrote it) · Jiaming He (Wrote it) · Zheming Zhang (Wrote it)
- What we read
- We have read only the description the publisher gives. No quotations until someone reads the source itself.
- Found by
- The Radar, run of November 6, 2026, via
arxiv-ai-safety-query(tier 1: the source’s own channel or an official record) - Cite this
- arXiv, “AUTEUR: Acoustic Plan Learning for Red-Teaming Text-to-Audio Models”, November 6, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-11-06-auteur-red-teaming-text-to-audio
Our notesLight lane: from the abstract in the triage excerpt, which appears complete.