Backdoored LLMs Pass Full-Precision Checks, Then Turn Malicious After Quantization
Post-training quantization is supposed to be a neutral optimization. This paper proves otherwise: adversarial fine-tuning can plant backdoors that pass every full-precision evaluation, then activate only after INT8 or 4-bit compression. In their tests, a supposedly clean translation model jumped from zero corruption at FP16 to 85% inversion once quantized. The implication: source-precision auditing is not enough, so the final deployed configuration must be included in behavioral certification for edge AI.