StealthRL preprint reports paraphrase attacks that evade multiple AI-text detectors

Researchers report in a preprint that StealthRL, a system trained to rewrite text against multiple AI-text detectors, sharply reduced detection performance and also evaded detector families excluded from training. They have publicly released the code and trained model checkpoint. The system uses reinforcement learning to train a rewriting policy against an ensemble of detectors while enforcing semantic fidelity. Across RoBERTa, Fast-DetectGPT, Binoculars, and MAGE, the mean true-positive rate at a 1% false-positive rate fell to 0.024. Quality evaluations found readable prose and nontrivial semantic overlap, which does not establish that every rewrite fully preserved meaning.