Speaker-Activated Backdoors Compromise Speech Models During Fine-Tuning

# Speaker-Activated Backdoors Compromise Speech Models During Fine-Tuning

https://gcdradar.vercel.app/incidents/speaker-activated-backdoors-compromise-speech-models-during · AIR-2026-1442 · closed 2026-11-04

- Lane: Research
- Severity: 1 (Research / proof of concept)
- Fix status: Fix status unknown
- First reported: 2026-10-21
- AI role: Attack on AI

## Summary

Researchers introduced speaker-activated backdoors that embed malicious triggers into speech models during fine-tuning using a target speaker's natural voice. The method succeeds without requiring access to inference-time inputs.

## What to do

No action recorded; read the sources below and the vendor's advisory.

## Latest updates

- 2026-10-21: You Are the Trigger: Speaker-Activated Backdoors (https://arxiv.org/abs/2610.27210)

## Sources

- 2026-10-21 arXiv: You Are the Trigger: Speaker-Activated Backdoors https://arxiv.org/abs/2610.27210

Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.