Widening the Gap: First Quantization-Conditioned Attack That Breaks GPTQ, AWQ, and GGUF I-Quants

arXiv·high signal

Researchers introduce an outlier injection attack that exploits how large outliers force surrounding weights to zero during quantization. The technique crafts models that appear benign at full precision but exhibit malicious behavior once quantized — and for the first time consistently compromises advanced methods including GPTQ, AWQ, and GGUF integer quants, where prior attacks failed. This expands the threat model for anyone deploying quantized open-weight models from community hubs.