FoeGlass: In-Context Learning Is Enough to Red-Team Audio Deepfake Detectors

arXiv 2606.05101·high signal

Dehdashtian et al. show that simple in-context learning suffices to red-team audio deepfake detection (ADD) models that guard against malicious TTS. The work exposes how easily ADD systems can be probed and evaded without elaborate attack infrastructure. A warning for builders deploying voice-authentication or deepfake-detection safeguards.