
claude-awm
Evades LLM text watermarks by injecting Unicode variation selectors; includes the SynthID generator, mean-g detector, normalization defenses, and…
adversarial-attackai-securitymachine-learning+2
61

Evades LLM text watermarks by injecting Unicode variation selectors; includes the SynthID generator, mean-g detector, normalization defenses, and…

Two-stage prompt-injection and jailbreak detector: regex gates plus a quantised DeBERTa-v3 ONNX classifier, with image, document, and audio support.…

StealthRL: RL framework for adversarially paraphrasing AI text to stress-test detector robustness.