An international group of researchers found that simply rerecording deepfake audio with natural acoustics in the background allows it to bypass detection models at a higher-than-expected rate.